Unique Vulnerabilities and Attacks on Cellular Data Packet Services By DENYS MA B.S. Computer Science and Engineering. (University of California, Davis) 2004

THESIS Submitted in partial satisfaction of the requirements for the degree of MASTER OF SCIENCE in Computer Science in the OFFICE OF GRADUATE STUDIES of the UNIVERSITY OF CALIFORNIA DAVIS

Approved:

Assistant Professor Hao Chen(Chair)

Professor Karl Levitt

Assistant Professor Xin Liu

Committee in Charge 2007

–i–

Unique Vulnerabilities and Attacks on Cellular Data Packet Services

Copyright 2007 by Denys Ma

ii

Contents 1 Introduction 1.1 Contributions of this Thesis to the Field . . . . . . . . . . . . . . . . . . . . . . .

1 2

2 Related Works 2.1 Cryptography . . . . . . . . . . . . 2.2 Cloning and Fraud . . . . . . . . . 2.3 Denial of Service . . . . . . . . . . 2.4 Spam and Phishing . . . . . . . . . 2.5 Worms . . . . . . . . . . . . . . . . 2.6 3G scheduling and network security

. . . . . .

3 3 4 4 5 5 6

. . . . . . . . . . . . . . . . . . . . .

7 7 8 8 8 9 10 12 12 13 13 13 13 15 15 16 16 17 17 17 18 19

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

3 Sleep Deprivation Attack 3.1 Background overview . . . . . . . . . . . . . . . . . . . . 3.1.1 GSM . . . . . . . . . . . . . . . . . . . . . . . . Location update . . . . . . . . . . . . . . . . . . . Paging . . . . . . . . . . . . . . . . . . . . . . . 3.1.2 GPRS . . . . . . . . . . . . . . . . . . . . . . . . 3.1.3 MMS . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Attacks . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.1 MMS security analysis . . . . . . . . . . . . . . . Unencrypted and unauthenticated MMS messages Unauthenticated MMS R/S . . . . . . . . . . . . . Critical phone information disclosure . . . . . . . 3.2.2 Attack implementation . . . . . . . . . . . . . . . Building target hit-list . . . . . . . . . . . . . . . Draining batteries . . . . . . . . . . . . . . . . . . Theoretical impact . . . . . . . . . . . . . . . . . 3.2.3 Attack experiment results . . . . . . . . . . . . . 3.2.4 Attack improvement . . . . . . . . . . . . . . . . Attack using TCP ACK packets . . . . . . . . . . Attack using packets with maximum-sized payload NAT and firewall . . . . . . . . . . . . . . . . . . 3.3 Mitigation strategies . . . . . . . . . . . . . . . . . . . .

iii

. . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . . . . . . . . . . . . . . . . .

3.3.1 3.3.2

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

19 20 20 21 22 24 24 25 26

4 Scheduler Attack 4.1 Attack overview . . . . . . . . . . . . . . . . . 4.1.1 3G data networks . . . . . . . . . . . . Opportunistic scheduling . . . . . . . . Handoff . . . . . . . . . . . . . . . . . 4.1.2 Overview of attacks . . . . . . . . . . 4.2 Attack analysis . . . . . . . . . . . . . . . . . 4.2.1 Attack within a single cell . . . . . . . Single attacker . . . . . . . . . . . . . Multiple attackers . . . . . . . . . . . . Simulation . . . . . . . . . . . . . . . 4.2.2 Attack from two cells . . . . . . . . . . Initial average throughput . . . . . . . Simulations . . . . . . . . . . . . . . . 4.2.3 Attack without knowing victims’ CQIs 4.3 Attack impact . . . . . . . . . . . . . . . . . . 4.4 Possible defense strategies . . . . . . . . . . . 4.4.1 Attack detection . . . . . . . . . . . . 4.4.2 Attack prevention . . . . . . . . . . . . 4.5 Conclusion . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

27 27 28 28 29 29 30 31 31 33 33 35 35 36 37 40 41 41 42 44

3.4

MMS Protocol Modification . . . . Adaptive PDP Context Management Motivation . . . . . . . . . . . . . Design Principle . . . . . . . . . . Strategy overview . . . . . . . . . . Specification Modification . . . . . Analytical Analysis . . . . . . . . . Implementation Details . . . . . . . Conclusion . . . . . . . . . . . . . . . . .

. . . . . . . . .

5 Summary and Conclusion

49

Bibliography

51

iv

Acknowledgements I want to give utmost gratitude to Professor Hao Chen for his most valuable advises, and guidance for not only this thesis, but also as a graduate student. This work would not exist without his insights and dedicated work. My gratitude also goes to Professor Karl Levitt for his help and advises in every step of my graduate life. He encouraged and supported me throughout the years I’ve been in Davis. I would like to thank everyone who contributed to this thesis. In particular, most credits to Radmilo Racic for his extremely valuable contributions. He has worked on this work in every aspect and help me through difficult problems. Also, my thanks to Professor Xin Liu for her contributions to this thesis. Many thanks are due to my friends for all the support and advises. My thanks to Dr. Jeff Rowe and Professor Felix Wu for all the advises on my research, Senthil Cheetancheri for his insights and discussions on worms and my research, and Allen Ting for his encouragement and support on my efforts. My gratitude to Carol Lin for her endless encouragements and support, even in difficult times. She believed in me even through periods of uncertainty, and gave me courage to proceed. Finally, I dedicate this thesis to my parents who, through hardship, provided me a chance to learn and discover as I wish. They have placed my needs over everything else to support me throughout my life. They have also shaped me into the person I am today through valuable wisdom and guidance.

v

Abstract As cellular data services and applications are being widely deployed, they become attractive targets for attackers, who could exploit unique vulnerabilities in cellular networks, mobile devices, and the interaction between cellular data networks and the Internet. Furthermore, mobile devices, often times considered to be part of the cellular network’s trusted computing base (TCB), are becoming more vulnerable to attacks. This thesis presents several vulnerabilities on the cellular data packet services and its applications, and present two particular denial of service attacks. First, we demonstrate an attack, which surreptitiously drains mobile devices’ battery power up to 22 times faster and therefore could render these devices useless before the end of business hours. This attack targets a unique resource bottleneck in mobile devices (the battery power) by exploiting an insecure cellular data service (MMS) and the insecure interaction between cellular data networks and the Internet (PDP context retention and the paging channel). Second, we propose a series of attacks on 3G cellular packet services that exploit the unverified channel condition reports from mobile devices to their base stations, and user-initiated handoffs. Our simulations show that only five rogue devices per cell can use up over 90% of the network resource, and thus induce and perpetuate 2.1s end-to-end inter-packet transmission delay for every user in the cell. This thesis also presents several mitigation strategies to defend against not only the two aforementioned attacks, but also similar attacks of these type.

vi

CHAPTER 1. INTRODUCTION

1

Chapter 1

Introduction Cellular networks are part of our critical information infrastructure. Cellular networks are also widely deployed, with more than 194 million subscribers covering over 65% of the US population. [1] As mobile devices become more powerful, cellular companies are rapidly deploying broadband data services, such as High-Speed Downlink Packet Access (HSDPA) and EvolutionData Optimized (EV-DO) as well as new applications, such as Multimedia Messaging Service (MMS), Unlicensed Mobile Access (enabling network-to-network mobile agent migration), i-Mode (providing fast, packet-based communication by eliminating the traditional WAP gateway), and Wifi Voice-over-IP (enabling affordable, realtime voice communication). Furthermore, cellular networks are pushing more network functions into mobile devices and grant them more trust. In some situations, they even consider mobile devices as part of the Trusted Computing Base (TCB). While these new services and applications enhance mobile computing experience, they also introduce serious security concerns. Besides launching typical Internet attacks — such as denial of service (DoS), malware, spamming and phishing — against mobile devices, an attacker can exploit emerging vulnerabilities in cellular networks, mobile devices, and the interaction between cellular data networks and the Internet. Emerging vulnerabilities in cellular networks, however are not thoroughly studied, both by the security community or service providers; since the cellular community are focused on information security rather than network security. We argue that network vulnerabilities can cause havoc in cellular networks, in particular, both current and future data services. Therefore, this thesis presents several emerging vulnerabilities in cellular data networks and two particular denialof-service attacks exploiting these vulnerabilities that can cause devastating affects. These attacks would be devastating not only in critical situations, such as disasters, but also for industries relying

CHAPTER 1. INTRODUCTION

2

on mobile communications. For example, professions like real estate agents and brokers rely on the ability to perform on-the-spot credit reports or provide instant quotes. Similarly, occupations such as network system administrators trust their cellular handset’s availability in order to be reached. The first attack, exploiting vulnerabilities in MMS and General Packet Radio Service (GPRS) in GSM, targets mobile devices’ battery power. The adversary is able to drain a mobile phone’s battery stealthily in 7 hours from the Internet. The second attack, exploiting vulnerabilities in 3G and 3.5G data packet services and their opportunistic scheduler, demonstrate that malicious mobile devices can usurp time slots at the expense of honest users, hence denying them network access.For example, we show that only one attacker per cell that has 50 users can occupy as much as 89% of the all the scheduling slots indefinitely. Similarly, five attackers per cell can cause and perpetuate 2.1s end-to-end inter-packet transmission delay for every victim user in the cell, thus rendering many services useless. This thesis proceeds by presenting an overview of the related works in cellular network security in chapter 2. Chapter 3 presents the first attack and the mitigation strategies that can defend against it. Chapter 4 presents the second attack, with the possible defense mechanisms. Finally, chapter 5 concludes this thesis.

1.1 Contributions of this Thesis to the Field This thesis makes the following contribution: • We identifies vulnerabilities in 2.5G and 3G data services and applications that relies on these services, in particular, MMS, GPRS, EV-DO, HSDPA, and the Proportional Fair (PF) scheduler. • We implemented an attack to surreptitiously drain a phone’s battery up to 22 times faster than normal, illustrating two key vulnerable components in the current cellular data networks. • We propose a series of attacks on opportunistic scheduling in 3G data networks, analyze these attacks mathematically, and explore the effectiveness of these attacks under different network configurations. Our simulations show that these attacks would devastate the network by rendering many services useless. • We propose approaches to mitigate or eliminate the impact of each attack.

CHAPTER 2. RELATED WORKS

3

Chapter 2

Related Works In recent years, significant amount of research efforts have been focused on security requirements and threat model evaluation on current and emerging cellular technologies, including GSM [2–4], GPRS [5–8], CDMA [9], SMS [10], MMS [11], and EVDO [12–14]. These works identify the following key security requirements in cellular networks: subscriber confidentiality, authentication, privacy, cloning prevention, integrity of information as well as billing, fraud detection, and safe key management. These works also address security threats such as eavesdropping, impersonation of a user and network, denial of service, man-in-the-middle attacks, hijacking services, and compromising authentication vectors. Apropos, researchers evaluated the risk levels of each of these threats as well. Our work is complementary to these previous efforts to secure cellular networks. In fact, we focus in two new directions: the end user devices (i.e., power-depletion attack and defense) and the security interactions between different cellular applications (i.e., the merging of cellular network and the Internet). In this chapter, we present an overview of the current research efforts in cellular networks.

2.1 Cryptography Extensive research has been conducted on the cryptography technologies [15–17]. For instance, studies like [15, 16] suggest the use of a PKI scheme in the GSM/UMTS network while [17] proposes the use of a SIM card for authentication and payment of web services by mobile users. Grecas and colleagues propose introducing public-private key pairs for transactions between the VLR-HLR as well as MS-VLR. Lo and colleagues, on the other hand, propose the use of PKI and stream ciphers for authentication and message encryption/decryption, respectively. They both point

CHAPTER 2. RELATED WORKS

4

out that the nature of the services constituting the PKI renders telecommunications operators prime candidates for the PKI implementation. Furthermore, MacDonald and colleagues [17] are convinced that SIM card can be at the center of an authentication and payment platform for consumption of web services by mobile users. Cryptographic solutions, while efficiently and elegantly mitigating some principal concerns in cellular networks, cannot defend against some unique threats to end users, such as a DoS attack and resource starvation attacks. Our work complements the existing cryptography mechanisms in order to alleviate additional non-conventional threats unique to emerging cellular data technologies and applications.

2.2 Cloning and Fraud Significant research has been done on mobile device cloning and the associated frauds [18]. In complementary to cryptographic solutions, schemes are developed to defend against cloning and fraud, such as device and user fingerprinting [19], mobility pattern recognition [20], and usage pattern recognition [21, 22]. These research studies propose new security mechanisms strictly for cellular networks. However, most studies stipulate fundamental changes in either architecture or end user equipment. In order to minimize disturbance of current implementation of cellular networks, our research will focus on utilizing existing security mechanisms to mitigate new attacks that were not discovered or considered.

2.3 Denial of Service Denial of Service attacks executed on 2G/2.5G networks also attracted a lot of attention, because resources in cellular networks are much more limited than on the Internet. In particular, control channels are in danger due to its narrow bandwidth.. Agarwal et al. [23] conducted a capacity analysis of shared control channels used for SMS delivery. They concluded that increasing volume and message sizes can significantly affect network performance. Then,, Enck et al [24] presented a denial-of-service attack by sending a sufficient number of SMS messages per second to a range of cellular phones in the same area. An attacker would need only a single computer with a broadband network access in order to disrupt a network in a major city by saturating control channels shared between voice calls and SMSs. Traynor el at.

[25] follows up on this work by simulating the

attack outlined in [24] using a highly accurate GSM simulator, and presented several mitigation strategies with supporting simulations. Additionally, [26] warns that paging channel is another

CHAPTER 2. RELATED WORKS

5

scarce resource that an attacker on the Internet can overwhelm and cause a DoS attack. Finally, Martin el at. [27] discussed the possibilty of a denial of service attack on mobile devices such as laptops and PDAs. They outlined three different types of battery draining attacks and presented experiments to demonstrate the affects of such attack. Nash el at. [28] follows up on the work by presenting a host-based intrusion detection system to detect battery draining attacks. Our work, inspired by these previous works, extends previous findings and presents additional vulnerabilities both in current and future cellular data services.

2.4 Spam and Phishing In addition to DoS attacks, spam is another well-known problem in the SMS network [10]. Network providers allow email and web-based interfaces to send SMS messages to individual or multiple handsets directly. Spammers can also employ phishing [29] to trick users into divulging private personal information. SMS-based phishing has already been discovered in a small German cellular provider [30], where users are tricked into sending a reply SMS to a value-added service’s SMS number, charging a small fee per user. Our first attack in Chapter 3 of building a hit-list of phone IP addresses and model information was inspired by phishing; however, our approach does not need the user’s participation or even attention, because such information is reported to our server automatically by most phones.

2.5 Worms Computer worms that target cellular networks have also appeared in recent years. Timifonica worm [31] spreads itself via email attachments. Upon infection, a computer sends SMS messages to random cell phone numbers belonging to a service provider, Movistar, and thus attempts to cause a DoS attack. A proof of concept worm was developed in early 2005 demonstrating the effects of a worm outbreak on cellular phone platforms. The Cabir [32] worm, spreading via Bluetooth on Nokia series 60 handsets running Symbian OS, changes the operating system and searches for other handsets to infect. An epidemic worm spreading model in mobile environments was proposed by Mickens et al. [33]. Our work is an extension to these previous works. Using a hitlist of phone numbers, IP addresses, and model information gathered in our attack described in Chapter 3, worm designers could write better worms by tailoring to different platforms.

CHAPTER 2. RELATED WORKS

6

2.6 3G scheduling and network security Significant amount of research has been conducted on efficient resource sharing in cellular networks. In particular, opportunistic scheduling algorithms have been studied extensively [34– 36]. However, the existing work focuses on improving system performance under various system constraints and requirements. For example, Choi et al [37] study the effects of Proportional Fair (PF) scheduler on TCP performance. They conclude that TCp’s minimum RTO is too short and it leads to unnecessary timeouts under the PF’s scheduling policy. Assaad et al. [38] report the effects of TCP on HSDPA operation and confirm that the lower the congestion rate of TCP, the higher the application bit rate is. Their results show that the effects of TCP on application performance are much higher than on the system capacity due to the use of the high speed shared channels. Andrews [39] also considered the PF scheduler suggested in the High Data Rate (HDR) data system and shows that the PF scheduler is unstable under certain conditions. Andrews defines stability as the ability to keep each user’s queue bounded. Using simulations and models, Andrews describes six different versions of PF scheduler and shows that all of them are unstable. Finally, Bu et al. [40] studied PF scheduler in multiple cells and propose a central PF scheduler to increase fairness. In contrast to our work discussed in Chapter 4, these studies does not consider potential threats of malicious users and the corresponding effects on the schedulers used in wireless systems. Initial studies on network security in 3G networks has also been published in recent years [41–43], outlining possible threats in the cellular network. Particularly, Sridharan et al. [44] model the uplink channel from mobile devices and the base station in EV-DO and suggest that malicious users can modify their power transmission level and cause interference for honest users. Our work in Chapter 4 differs from their work by concentrating on the downlink, since in 3G networks, downlink bandwidth is much higher than uplink. Furthermore, these studies do not provide an actual attack, but only outline possible threats against 3G networks.

CHAPTER 3. SLEEP DEPRIVATION ATTACK

7

Chapter 3

Exploiting MMS Vulnerabilities to Stealthily Exhaust Mobile Phone’s Battery In this chapter, we present an attack that exploits vulnerabilities in MMS (a cellular data service), PDP context retention in GPRS (interactions between the Internet and cellular data networks), and the paging channel. Furthermore, this attack has unique features that (1) it is clandestine – victim mobile users will not notice when their batteries are being drained; (2) it is not limited to certain mobile device hardware or software; and (3) it targets individual mobile devices rather than the network, an attack that is often harder to detect and defend effectively by network operators. We implemented this attack in two stages. In the first stage, we were able to build a fairly accurate ”hit-list” of all the users with an active Internet connection by taking advantage of the insecure MMS protocol. In the second stage, we exploit the PDP context retention to surreptitiously drain a phone’s battery up to 22 times faster than normal. This attack illustrates two key vulnerable components in the cellular data network, and we will propose mitigating strategies for securing these components.

3.1 Background overview To help understand the vulnerabilities and attacks that we discovered, we present an overview of the relevant components in cellular networks: GSM, GPRS and MMS.

CHAPTER 3. SLEEP DEPRIVATION ATTACK

8

3.1.1 GSM The key elements in GSM are: the Base Station Subsystem (BSS), which includes the Base Transceiver Station (BTS) and the Base Station Controller (BSC), and Mobile Switching Center (MSC) which is the core of the Network Sub System (NSS). Additionally, these GSM elements utilize databases like Home Location Register (HLR) and Visitor Location Register (VLR) for storing users’ home as well as roaming information, respectively. BTS provides the means to transmit and receive radio signals as well as encrypt and decrypt communication with the BSC. BSC provides network intelligence by allocating radio channels, controlling inter-BTS hand-offs and, most importantly, serving as a gateway to the MSC. MSC, on the other hand, sets up circuit-switched communications, takes care of mobility management and manages other databases. A cellular network needs to keep track of the location of each Mobile Station (MS1 ) in order to deliver calls and data to the correct destination reliably. Typically, the network utilizes an event-based mechanism to collect mobile device’s location. Events such as powering up, shutting down, and crossing into another location area are events that trigger the location update procedure. A cellular network is partitioned into cells serviced by BTSs. Cells are then grouped together to optimize signaling and to facilitate tracking of mobile phones within the network. Each group, managed by one BSC, is identified by a location area code broadcast by each BTS at regular intervals. Two fundamental operations within the location area are location update and paging. Location update The MS sends location update messages to its current BTS periodically in order to route all incoming calls or data appropriately. If the MS sends updates seldom, its location is unknown and the MS must be paged for each downlink packet (or call), thus degrading the quality of service. If, on the other hand, the MS sends frequent updates and its location is known, then data packets can be delivered without any additional paging delay. Paging To minimize the amount of updates, preserve MS’s battery, and minimize bandwidth utilization, the network will page the MS over the Paging Channel (PCH) to determine its location. In 1 MS

and phone will be used interchangeably.

CHAPTER 3. SLEEP DEPRIVATION ATTACK

9

MSC VLR

HLR

SGSN

Internet BSS MS

GGSN

SGSN on another PLMN

Figure 3.1: GPRS infrastructure other words, PCH is used for communication from BTS to MS when MS is not assigned a traffic channel; that is, the MS’s location is unknown or out of date. The paging bandwidth burden is relatively small in small location areas - less than 1% of the bandwidth allocated for voice channels. On the other hand, in an area with a large number (over 1000) of cells per location area, the paging bandwidth burden could be considerably higher. [45]

3.1.2 GPRS GPRS [46] is integrated into the existing GSM infrastructure with a new class of network nodes called GPRS Support Nodes (GSNs). GSNs are responsible for the delivery and routing of data packets to and from the mobile network. There are two types of GSNs: Serving GPRS Support Node (SGNS) and Gateway GPRS Support Node (GGSN). SGSN is responsible for transferring and routing of data packets, mobility management, logical link control, authentication and billing services within its service area. GGSN acts as an interface between the GPRS backbone and external packet networks (primarily the Internet). Its primary function is to convert GPRS packets coming from the SGSN to IP packets and vice versa. An illustration of GPRS is shown in Figure 3.1. Before an MS can utilize GPRS services, it must register with an SGSN so all packets can be routed through it. During this procedure, called GPRS attach, a PDP (Packet Data Protocol) context is created. In particular, SGSN checks if the user is authorized, copies the user profile from the HLR to itself, assigns a Packet Temporary Mobile Subscriber Identity (P-TMSI)2 , maps it to an IP address, and assigns a GGSN that will serve as the gateway to the Internet. The PDP context, 2 The

reasoning is to minimize use of IMSI (International Mobile Subscriber Identity) for security purposes.

CHAPTER 3. SLEEP DEPRIVATION ATTACK

10

composed of the above mentioned information, is stored at the SGSN. GPRS detach, on the other hand, disconnects the MS from the GPRS network and deactivates the PDP context. Location areas have been proven to be efficient in voice networks; however, the bursty nature of data traffic increases the number of paging messages per phone in each location area. Therefore, each location area is further subdivided into routing areas used by GPRS to decrease the penalty for locating an MS. GPRS phones utilize IDLE, STANDBY and READY states in increasing order of battery consumption. When an MS is in the READY state, SGSN is aware of the MS’s location. In particular, the MS performs frequent location updates to provide the network with the actual cell ID so that no paging is necessary. When in the READY state, the MS can send and receive data. Furthermore, it will stay in the READY state until READY timer expires, at which it will transition to the STANDBY state. While in the STANDBY state, the MS has established the PDP context and it can receive calls or data. However, its location updates are more coarse, in the sense that it informs the SGSN of only routing area changes, but not cell changes. If SGSN needs to deliver data to the MS while the MS is in the STANDBY state, SGSN will send a page request in the routing area where the MS is located. When MS responds to the page, it will transition to the READY state. IDLE state is the lowest battery consumption state, in which the SGSN is not aware of the MS’s location. The MS can transition out of IDLE state only if it performs a GPRS attach procedure. Alternatively, an MS could initiate a GPRS detach procedure to transition to the IDLE state. Figure 3.2 shows the state machine of the GPRS MS. Upon completion of the communication, the MS will go into a STANDBY mode. The PDP context, on the other hand, will remain allocated to the MS. We conducted experiments to discover how long each handset retained its assigned PDP context and IP address. We found that addresses seemed to be relinquished in as short as 15 minutes to as long as several hours. The reason for not deactivating a PDP context is simple: a cellphone can be unavailable for a period of time due to radio link failure; deactivating and activating a new context would imply that the phone would need to recreate all TCP sessions, possibly restarting applications and requiring the user to re-enter all the passwords.

3.1.3 MMS MMS has become a very popular cellular message service. The MMS architecture spans both the cellular network and the Internet and uses technologies in both networks, such as WAP, SMTP, and HTTP.

CHAPTER 3. SLEEP DEPRIVATION ATTACK

11

PDP CONTEXT INACTIVE

IDLE

GPRS Detach

GPRS Attach

READY

STANDBY timer expired

Data transmit or receive

READY timer expired Force to STANDBY

STANDBY PDP CONTEXT ACTIVE

Figure 3.2: The GPRS mobile station state machine The MMS architecture consists mainly of the MMS Relay/Server (MMS R/S) and user agents. Several optional entities of the architecture – the billing server, the Home Location Register, and the User Database — may exist inside or outside MMS R/S. Figure 3.3 shows an overview of the MMS architecture. The MMS R/S is responsible for all of the transactions of MMS. When a user transmits an email or an MMS message, the mobile phone formats these messages in Synchronized Multimedia Integration Language (SMIL) [47]. The MMS R/S translates (transcodes) the message to either email or different MMS formats depending on the provider. The message is then sent to the destination SMTP mail server or the destination MMS R/S using SMTP. Upon receiving the message, the destination MMS R/S then stores the message in the user’s buffer while sending a notification message to the user via a SMS or WAP push message. The notification message contains the location of the message, usually specified as an HTTP address. User can configure their mobile phones either to automatically download the message upon receiving the notification or to manually download the message themselves.

CHAPTER 3. SLEEP DEPRIVATION ATTACK

HLR

12

User dB

Billing Server

MM8

MM9

User Agent

Email Client

MM5

SMTP

MM1

MMS R/S

Wireless Network

Internet MM1

MM4

Figure 3.3: MMS Infrastructure

3.2 Attacks In this section, we present our findings on attacking the cellular network. We first investigated the MMS protocol and discovered several vulnerabilities through which we leveraged into the heavily protected cellular network. Then, by exploiting these vulnerabilities, we implemented a proof-of-concept attack on a scarce resource – the battery power – of mobile devices. The attack is stealthy, as it is noticeable to neither mobile users nor network operators. Our experiments demonstrate that unique threats against cellular networks and mobile devices exist and are exploitable. Finally, we discuss how to make this attack even more effective.

3.2.1 MMS security analysis To test how cellular providers implement MMS and gain insight into their interface designs, we setup our own MMS R/S, based on an open-source project [48]. We discovered several vulnerabilities that a wily attacker could exploit, as described in the following sections.

CHAPTER 3. SLEEP DEPRIVATION ATTACK

13

Unencrypted and unauthenticated MMS messages We confirmed that MMS messages and MMS notification messages, composed of headers and content sections, were sent in plain-text. In addition to the SMIL headers, the packet also included an HTTP POST header containing the source and destination IP address, the profile of the user agent, the content type and size, and the user agent name. Unauthenticated MMS R/S To mitigate the problem of unencrypted messages, cellular providers hide their own MMS R/S’s IP addresses in the phones, hoping that cellular users cannot read or overwrite them. Unsurprisingly, we discovered that this attempt at security by obscurity is broken. In order to inspect the MMS message raw format, we modified a phone’s firmware to route all MMS messages through our MMS R/S. The MMS R/S setting is well hidden in our phone’s firmware, which suggests that providers do not intend to allow users to modify the setting. After modifying the MMS R/S entry in our phone, we discovered that the phone had no security mechanism to alert the new, unauthorized MMS R/S. Furthermore, MSs also do not authenticate MMS notification messages and MMS messages sent from the network. MSs will accept any MMS messages as long as the format is correct. Consequently, we were able to send unlimited MMS messages for free, without alarming the cellular provider. Critical phone information disclosure We discovered that handsets include pertinent user agent platform information whenever they communicate over HTTP. Accordingly, we set up a web server running ethereal to capture HTTP requests from various handsets on different networks. We found that every phone disclosed either its full profile or information that included one or more of the following: hardware platform description, display capabilities, and the current and compatible software. An attacker could write a script that extracts the model number of each handset very easily.

3.2.2 Attack implementation Based on our MMS security evaluation, we implemented a battery draining attack utilizing a hit-list built using superfluous but pertinent information disclosed during MMS exchanges. Figure 3.5 illustrates the attack.

CHAPTER 3. SLEEP DEPRIVATION ATTACK

14

Figure 3.4: Ethereal reconstruction of an MMS message captured by our MMS R/S. The message is transported in clear text. Various fields such as the server, the sender’s phone number and phone model are exposed and could be collected in a the hit list.

MMS Server

(2 n)

(2 1 )

(3 n)

Attacker

.

(3 1)

Victim n

Victim 1

(1 1 )

(1 n)

Figure 3.5: A two-step attack on cellular devices. In Step 1, the attacker builds a hit list using MMS message notifications (Messages (1)s), and captures information about mobile users from the HTTP requests from mobile users (Messages (2)s). In step 2, the attacker drains the batteries of cellular devices on the hit-list surreptitiously by sending UDP packets (Messages (3)s) periodically to the cellular devices.

CHAPTER 3. SLEEP DEPRIVATION ATTACK

15

Building target hit-list To launch effective, large scale attacks, an attacker needs to build a hit-list that contains important information about the network and end users. One way to obtain such information is by asking the mobile phones. An attacker can send MMS notification messages, whose content address is at a malicious web server, to numerous recipients. The target phone numbers can be generated automatically using known area codes and prefixes for cellular phone numbers. The MMS notification messages can be sent using SMS or WAP push. There are many free SMS messaging websites, including those offered by cellular providers. Once MMS notification messages are sent, the attacker waits for HTTP request messages at his web server, which has stated its location in the MMS notification message. Since many cell phones are configured to download MMS messages automatically upon receiving notification, they will make HTTP requests to the attacker’s web server. The HTTP requests often contain the profiles and IP addresses of the phones, and even file extensions that the phones are able to process. By sending a slightly different URL to each phone, the attacker can build a hit list that maps each phone number to a profile of its cellular device. More importantly, the phone’s response to the MMS notification message activates a PDP context, making our attack easy and simple to execute even in the presence of NAT and firewalls. Draining batteries Using the hit-list generated from MMS notification messages, an attacker can target the cellular network and cellular devices more precisely and effectively. Apropos, we implemented a battery draining attack that focuses on the end hosts instead of the network. We implemented our attack using UDP packets (we will explain an improved technique later.) The key to maximizing a cell phone’s battery life is to use its transceiver sparingly. In fact, when a cellular phone is turned on, its transceiver is active less than 3% of the time. As a reference, in wireless sensor nodes, transmitting one bit of information consumes 1500 to 2700 times as much energy as executing one instruction [49]. Thus, if a packet is sent to a phone, the SGSN will deliver the packet if the phone’s location is known, or attempt to locate the phone by sending a page request to it. However, since cellular phones spend most of their time in the STANDBY mode (or other dormant modes), the page on the paging channel will awaken the phone to the READY state and force it to perform a location update. The sine qua non of this attack is to keep the phone in the

CHAPTER 3. SLEEP DEPRIVATION ATTACK

16

READY state (high battery consumption), therefore disabling its ability to preserve battery life, or to let the phone temporarily go into the STANDBY state only to be immediately awakened with a page and forced to perform a location update; both of these actions consume much energy. Theoretical impact To investigate the severity of the aforementioned attack, we estimate the damage that an attacker with a home DSL Internet connection can inflict. A typical DSL upload speed ranges from 256kbps to 416kbps. We use the medium speed, B = 384kbps, for the upload bandwidth as an estimate. Each UDP packet consists of a character in the data segment, which might be padded to 4 bytes depending on the provider’s DSL modem. The UDP packet header has 8 bytes, and the IP header has 20 bytes. In the pessimistic estimate where our data is padded, the total size of the packet is S = 32 bytes. Therefore, the maximum number of UDP packets per second that an attacker may send is (B/8)/S = 1500. To attack a phone effectively, an attacker must send one UDP packet to the phone every T seconds. In this case, the maximum number of phones that the attacker can attack simultaneously is (B/8) ∗ T /S. We estimated the time T by trial and error using different test configurations. For our experiment, we chose 3.75 seconds for the GSM-based network and 5 seconds for the CDMAbased network. Using our equation, we calculated that an attacker can attack about 5625 phones using a standard ADSL line for a GSM-based network and around 7000 phones for a CDMA-based network.

3.2.3 Attack experiment results We successfully drained our test phones’ batteries considerably faster than our average usage. We conducted six test runs on a high-end Nokia smart phone and completely drained its battery in an average of 7 hours, instead of 156 hours in normal usage with bluetooth switched off most of the time. We also observed severe battery exhaustion in our Sony Ericsson test phone, where the battery was drained down to 20% within less than 7 hours without talking and with bluetooth switched off. If a phone is connected to the Internet continuously (for example, to use the instant messaging service), its battery life would be reduced much faster. To test this hypothesis, we attacked our Motorola test phone while connecting it to the Internet continuously. Our test completely drained its battery within 2 hours. Table 3.1 summarizes the results of our attack. We successfully conducted our attack on two major cellular service providers without

CHAPTER 3. SLEEP DEPRIVATION ATTACK Phone Nokia 6620 Sony-Ericsson T610 Motorola v710

Battery Life Without Attack Normal Use (hours) Standby (hours) 156 200 60 315 36 150

17 Battery Life Under Attack Normal Use (hours) Reduction 7 22.3:1 7 8.6:1 2 18.0:1

Table 3.1: Reduction of battery life due to our attack triggering any alarms. Our test machine’s IP was not blocked, our phones were fully operational after the attacks, and no notifications or warnings were sent to us regarding this issue. Moreover, during the attack the phone appeared to be operating normally and no additional Internet application was started, so the victim user would not notice the attack, until his/her battery died unexpectedly.

3.2.4 Attack improvement There are several optimizations that could be done to improve our attack. Currently, we empirically determined a fixed interval between each UDP packet by trial and error. However, by using Qualcomm’s CAIT software or knowing the implementation of a particular cellular network, we could obtain more accurate wait-time and thereby improve efficiency of our attacks. Also, knowing which IP addresses are vacant would increase the efficacy of our hit-list creation. We are currently in the midst of testing the following improvements to our attack. Attack using TCP ACK packets To force a phone to send as well as receive useless data, an attacker can periodically send TCP ACK packets to the phone’s IP address. In accordance with RFC793, if the connection is reset or in half-open state, the receiver of an out-of-order ACK packet will send an RST packet. If, on the other hand, the connection is open, the receiver of an out-of-order ACK packet will reply with an empty packet. Either way, an attacker will force a phone that implements a full TCP stack to receive as well as send packets, thereby exacerbating the power consumption. Attack using packets with maximum-sized payload In implementing our previous attack, we used UDP packets with no payload in order to maximize the number of UDP packets an attacker can send per computer. However, this is not the most efficient method of draining a cellular phone’s battery, since the whole packet must

CHAPTER 3. SLEEP DEPRIVATION ATTACK

18

be downloaded to the mobile phone before the phone can discard the packet. Therefore, with an accurate hit-list collected using MMS above, the attacker can sacrifice the number of targets per his/her computer to deliver an even more efficient attack using a maximum-sized payload. Using the original attack implemented with UDP, the attacker can send a maximum theoretical UDP data packet of 64Kb due to its 2 byte total length field. In the TCP variety of the attack, ACK messages ”piggyback” onto the existing payload with a maximum size of 1500 bytes. Besides causing additional unnecessary downloads for the mobile agent, the attack could possibly be even more efficient due to packet fragmentation. This exacerbates the attack so that the attacker would only need to send a single packet that becomes multiple packets at the mobile agent. NAT and firewall Through field experimentation, we have determined that most providers who utilize NAT also implement Network Address and Port Translation (NAPT.) NAPT provides dynamic (privateIP, privatePORT) to (publicIP, publicPORT) translation. For example, the inside interface tuple (10.0.0.5, 3000) could be mapped to the outside interface tuple (199.156.3.4, 6000). However, there are certain issues with network-wide NAT deployment. For example, it often hinders application deployment. Additionally, certain security protocols such as IPSec and Kerberos are affected – NAT changes the address in the IP header, causing loss of integrity. For these reasons, operators choose to implement NAT only on certain subnets affecting a selected customer base. In other words, most operators offer both private and public IP plans. It would seem that our attack could be mitigated with NAT and firewall placement. However, a very simple restriction to the attack could yield the same result. The crux of the change would be an observation that each inside IP address maps to a port on the outside interface because the publicIP is the public IP address of NAT system. Thus, targeting an inside IP address reduces to targeting a certain port of the outside interface. Since NAPT does address and port translation dynamically, the IP address and port mappings are only alive during active PDP contexts. Thus, the attack must be delivered within an active session window. Since phones automatically create an outbound connection to connect to a malicious HTTP server, the server itself must deliver the attack, thus prolonging the connection. The firewall would consider this connection valid as it is internally initiated over allowed ports, and NAT would continue the address and port translation for the duration of the attack.

CHAPTER 3. SLEEP DEPRIVATION ATTACK

19

3.3 Mitigation strategies Our attack uncovered two vulnerable components in cellular networks. • PDP Context is retained. We observed that a mobile user’s PDP context is kept alive even after the user has completed his/her data session. The PDP context may be kept active from 15 minutes to several hours, depending on the service provider. This active PDP context allowed us to send unwanted IP packets to the victim’s mobile phone to drain its battery. • Attack packets are not in any active session. Our attack periodically sends packets to mobile user without an active connection. A mobile user must initiate active connections before he receives data. Since the GGSN records the connection states, it can distinguish attack packets from normal packets that belong to active connections, unless the attacker can guess the correct sequence number, destination IP address and port number of an active connection. Based on these observations, we suggest the following mitigation stratgies on MMS and GPRS.

3.3.1 MMS Protocol Modification To mitigate threats against MMS, we propose a redesign by incorporating security mechanisms into the protocol. • Message and server authentication. To avoid man-in-the-middle attacks, we should authenticate MMS messages and R/Ss, using PKI for instance. • Information hiding at WAP gateway. WAP gateway should prevent outside web servers from obtaining critical information about mobile devices, such as their IP addresses, and hardware and software profiles. Since profiles are used only by the WAP gateway for converting web contents, the WAP gateway should filter out all but essential information about the user agent in HTTP requests. • MMS message filtering. Service providers typically hard-code their approved MMS R/S into mobile devices’ OS or firmware to prevent users from choosing alternative MMS R/Ss. However, sophisticated users can modify their OS or firmware to defeat this protection. A more reliable approach for service providers is to filter MMS messages, since all MMS packets must traverse the provider’s network. The filter can scan MMS message headers to ensure

CHAPTER 3. SLEEP DEPRIVATION ATTACK

20

that the destination IP address is one of the MMS R/S or accredited third party Value Added Service (VAS) providers. The filter should not be implemented at the WAP gateway, but rather at the SGSN or GGSN, since users can easily modify the phone’s settings and bypass the cellular provider’s WAP gateway.

3.3.2 Adaptive PDP Context Management In addition to protocol modification, we suggest a defense framework that could avoid the shortcomings of external firewalls and IDSs mentioned above by supplementing these protection mechanisms: • This defense mechanism can also serve as an event detector for IDSs already in place to monitor the internal network. • It must also be effective against insider attacks, where malicious users are connected using the cellular network instead of the Internet. • It should be designed with the goal of being non-intrusive so that it does not require ancillary network infrastructure; it should utilize existing GPRS mechanisms to provide an additional layer of protection. In the following section, we propose a novel defense mechanism implemented at the GGSN, Adaptive PDP Context Management (APM) designed to detect and mitigate previously mentioned attacks. Motivation Firewalls and IDSs are common mechanisms for defending against malicious behavior from the Internet, but they have several disadvantages: (1) firewalls and IDSs become the single point of failure, (2) they are external entities, and they usually do not protect against insider attacks, (3) they are not flexible enough to dynamically adapt to traffic conditions without system administrators – they require knowledgeable administration staff, (4) they are not suitable for monitoring peer-to-peer (such as Bluetooth) communication, and (5) they cannot protect against attacks exploiting insecure protocols whose action is seemingly valid – they either allow or deny a connection. Our defense framework attempts to avoid these downfalls of external firewalls and IDSs by supplementing these protection mechanisms in order to detect and mitigate attacks that could

CHAPTER 3. SLEEP DEPRIVATION ATTACK

21

stealthily bypass firewalls and IDSs. Our defense mechanism can also serve as an event detector for IDSs already in place in order to monitor the internal network. Our defense mechanism is also effective against insider attacks, where malicious users are connected using the cellular network instead of the Internet. Finally, APM is non-intrusive – it does not require ancillary network infrastructure as it utilizes existing GPRS mechanisms to provide an additional layer of protection. Using these two observations, we developed APM to detect and mitigate attacks on the GGSN. APM, not only can completely mitigate our battery draining attack, but also detect and mitigate other attacks exploiting the paging channel and PDP context, such as flooding attacks on the paging channel using packets from the Internet. Design Principle We designed APM with three goals in mind, • It should be implemented in the network core. • It should be transparent to mobile users. • It must be simple. Since our attack focuses on draining the battery of mobile users, the defense strategy should not exacerbate the attack by requiring additional processing from the mobile phone. If this was not the case, the defense mechanism itself could be utilized as a battery draining tool. Since the network core is assumed to have unlimited battery power, we must implement the defense mechanism at the core. Furthermore, it is almost impossible to implement any defense strategies on the mobile phone since cellular technology has already been widely deployed. Service providers cannot require all users to upgrade or update their hardware. Any type of defense strategy would be useless if users do not implement the mechanism. For instance, software patches are often useless against malware due to deployment issues. On the other hand, cellular providers can easily deploy defense strategies at the core, without user interaction. Our defense should also be transparent to each user. If our defense mechanism causes any inconvenient for mobile user, user will most likely complain to service providers. Usability is a main concern for mobile users since attacks on mobile phones are, at this time, unlikely and not wide spread. Furthermore, service providers will be less inclined to implement our strategy due to the inconvenience for users and the support cost to educate customers.

CHAPTER 3. SLEEP DEPRIVATION ATTACK

22

Packet

Outgoing

Yes

No Yes

PDP Context Exists

New Connection

No

No

Yes

Drop Packet Existing connection

No Yes

Yes

Transmit Packet

Transmit Packet

stateCount > 0

stateCount < max

Connection close

No

stateCount / 2

Yes stateCount * 2

Yes

No (1) Backoff (2) PDP Modification (3) Drop Packet

(1) stateCount– (2) Transmit Packet

Figure 3.6: Adaptive PDP Context management scheme Finally, our defense strategy should be as simple as possible due to the high workload of each GGSN. GGSN is responsible for providing an interface for millions of mobile phones. If our mechanism is computationally consuming, the attacker can exploit this vulnerability and in turn cause a DoS on the GGSN. Strategy overview For clearity, we present APM in both pseudo code shown below, and state diagram shown in Figure 3.6. APM is separated into three phases, the detection phase, exponential increase linear decrease (EILD) phase, and the recovery phase. APM(packet) 1

if packet is outgoing

2

then if packet initiates a new connection

3

then if statecount < statecountmax

4

then statecount ∗ 2

5

if packet ends a connection

6

then statecount/2

7

else if PDP context exist

8

then if packet does not belongs to existing connection

CHAPTER 3. SLEEP DEPRIVATION ATTACK

23

then if statecount > 0

9 10

then statecount − 1

11

else backoff

12

perform PDP Modification

13

drop packet

14 15

else drop packet The APM detection algorithm works in the following manner. For each packet, the algo-

rithm decides if it’s valid or not. The GGSN can accomplish this by using our second observation discussed in Section 5.1, a packet is not valid if it is incoming, and does not belong to any active connections. Since GGSN is stateful, and already keeps track of connection states, it can distinguish if a packet is valid or not by simply examining the header of each packet. However, to offload work from the GGSN, we also propose a modification to the PDP context. Currently, PDP context only contains the external address of each mobile device. Instead of simply storing the address, we can store (IPaddress, portnumber) tuple. A modified PDP context can have multiple address and port tuples. Whenever a mobile agent requests an outgoing connection, a tuple is assigned to it instead of just an address. Using this technique, we can easily distinguish between valid and non-valid incoming packets. To manage PDP context lifetime, we introduce a new variable along with the PDP context called stateCount. This counter serves as the time to live (TTL) time for each PDP context. The algorithm uses the stateCount variable in the following way: when GGSN receives an outgoing connection request or packet, a new tuple is assigned to the mobile phone, and the stateCount is doubled. If stateCount is 0, then we initialize it to 1. However, if GGSN receives an incoming non-valid packet, the stateCount is decremented by 1. When stateCount decreases to 0, GGSN can conclude that the phone is under attack and perform recovery. This phase is called exponential increase, linear decrease (EILD). By implementing EILD, our algorithm can withstand some amounts of false positive readings before raising an alert and entering the recovery phase. For example, it would be hard to distinguish between valid and malicous streaming traffic. Furthermore, many port scanners, worms, and other backscatter activities [50] are unavoidable on the Internet. Using EILD, we can avoid disrupting the user as much as possible before we enter the recovery phase. Finally, we note that PDP context can still be kept indefinitely, depending on service provider’s policy, as long as the mobile agent is not under attack.

CHAPTER 3. SLEEP DEPRIVATION ATTACK

24

The recovery phase is implemented when the stateCount decrements to 0. The recovery phase is implemented as follows: before disconnecting the user, we implement a random backoff wait period between (Cmin ,Cmax ). The wait period allows any existing connection to finish. After a random backoff waiting period, GGSN implements a gateway assisted PDP context modification, changing the external address in the PDP context. At this time, all connections currently still active will be dropped, thus preventing the attacker from reaching the mobile agent. Furthermore, only one extra message would be sent to the mobile agent notifying the modification, and one extra message would be sent from the mobile agent acknowledging the change. The mobile agent, after the recovery phase, can resume data connection and request outgoing connections as usual. Specification Modification Our defense strategy can be safely implemented in existing GPRS infrastructure without any violation to the specification [51]. In particular, GPRS specification states that user should be able to establish and deactivate GPRS service as requested. Our defense mechanism does not violate any of the specification stated. The specification does not clearly state any PDP context management schemes. In fact, the specification does not restrict when PDP context should be deactivated. However, the specification, under invocation and operation, states that, It shall be possible for a MS to be a GPRS service requester and service receiver. Our defense mechanism would violate this specification. However, we argue that cellular devices should not act as a server or any service receiver. In fact, most service providers in the US restricts mobile agent’s usage and does not allow any type of services to be active on any mobile users. Furthermore, the specification allows our battery draining attack, and many other attacks possible since it is allowed for an entity to activate the PDP context and communicate with mobile devices. We argue that such action should not be encouraged and protection against such exploitation should be straightly enforced. Analytical Analysis We now present an analytical analysis of our proposed defense strategy and provide a simplistic calculation of the maximum stateCount value which must be set in order for our defense to detect an attack.

CHAPTER 3. SLEEP DEPRIVATION ATTACK

25

We define the number of packets needed in order to mount a battery draining attack as follows:

n× Given n =

#packets s

60sec 60mins × × hhours = 602 nh. min hr

(3.1)

and h, the number of hours required to drain a phone’s battery (h ∼ 1n ) we can

calculate the upper bound on the number of outgoing connections that a cellular operator may set. Parameters n and h are network dependent so each operator would have to tailor them to their network. In order to detect this attack, our stateCount variable must not exceed 602 nh. Since we exponentially increase stateCount in the fashion of 2connectionCount , we calculate connectionCountmax as follows:

connectionCountmax = ⌊log2 602 nh⌋

(3.2)

And following our argument from above, the connectionCount should be calculated as follows:

connectionCount ≤ ⌊log2 602 nh⌋

(3.3)

For example, for n = 1packet and h ∼ 4 hours we notice that connectionCount = ⌊log2 14400⌋ = s 13. This means that the maximum number of connections each phone can make simultaneously in order to detect an attack that sends

1packet s

is 13 connections.

Note that this calculation provides a maximum for the connectionCount variable. Providers should set this variable limit to a much smaller number, in order to detect any type of attack much faster than this rate. Implementation Details As mentioned previously, our defense strategy is best implemented on the GGSN. Since service providers already perform some proprietary PDP management scheme, as tested empirically 3,

implementing our scheme would be very simple. Furthermore, as most of the functions needed 3 During

our battery draining experiments, the PDP context sometimes would detach even if the mobile phone is stationary. We notice that PDP context can be alive from 15 minutes to even days.

CHAPTER 3. SLEEP DEPRIVATION ATTACK

26

are already implemented, such as the gateway assisted PDP context modification function, there would not be any additional implementation work. Furthermore, our proposed extension on the PDP context would also be a simple modification. The implementation of the modified PDP context can be transparent to mobile devices, and the mapping can be done entirely at the GGSN. Since GGSNs are already stateful, a simple change in IP address assignment would not be difficult. Furthermore, our proposed modification to the PDP context would also provide a NAT like behavior, as each IP address can be assigned multiple times using different ports. We envision APM to be implemented as a ”plug-in” module, which should not be any longer than a couple of hundred lines of code. Since GGSNs are standardized within each service provider, a patch-like distribution can be easily deployed once the module has been fully tested on testbeds.

3.4 Conclusion In this chapter, we demonstrated an attack, such that is able to drain mobile devices’ battery power as much as 22 times faster. This attack proceeds in two stages. First, the attack exploits vulnerabilities in MMS to build a hit list of mobile devices. Then, the attack exploits PDP content retention and the paging channel to drain mobile devices’ battery power. We were able to drain batteries without alerting either the mobile user victims or the cellular network operators. Our analysis shows that an attacker would need only several home DSL Internet connections to mount a large scale attack against a large number of cellular phones. We identified key components in cellular networks that enable this attack and proposed corresponding mitigating solutions.

CHAPTER 4. SCHEDULER ATTACK

27

Chapter 4

Exploiting Opportunistic Scheduling in Cellular Packet Networks In this chapter, we study 3G and 3.5G networks and investigate the unwarranted trust granted to mobile devices and the ensuing vulnerabilities. These networks rely on schedulers to multiplex the spectrum efficiently. A commonly used scheduling algorithm is Proportional Fair (PF) [52, 53], which maximizes the product of the throughput delivered to all users. In this paper, we reveal vulnerabilities in the PF scheduler and demonstrate that malicious mobile devices can usurp time slots at the expense of honest users, hence denying them network access. For example, we show that only one attacker per cell that has 50 users can occupy as much as 89% of the all the scheduling slots indefinitely. Similarly, five attackers per cell can cause and perpetuate 2.1s end-toend inter-packet transmission delay for every victim user in the cell, thus rendering many services useless.

4.1 Attack overview Our attacks exploit vulnerabilities that result from unwarranted trust that the network grants to mobile devices. By reporting false channel conditions and initiating frequent handoffs, attackers can usurp the majority of downlink1 scheduling slots, causing intolerable delays to the victim users and rendering many network services virtually useless. We will give an overview of the vulnerable 3G data network technologies and our attacks. 1 From

the network to the mobile users.

CHAPTER 4. SCHEDULER ATTACK

28

4.1.1 3G data networks Cellular providers have developed two new data services, EV-DO and HSDPA, to provide broadband like downlink speed for emerging applications such as Voice-over-IP (VoIP), and streaming video and audio, without major network restructuring. In both services, the downlink utilizes time division multiplexing by dividing the channel in time slots, or Transmission Time Interval (TTI). (Note that T T I = 1.67ms for EV-DO and T T I = 2ms for HSDPA). The scheduler at each base station then selects a single user to transmit at each TTI. Both services rely on two main techniques to increase efficiency in the downlink direction: link adaptation and fast retransmission. Link adaptation utilizes base station’s processing power to collect quasi instantaneous downlink quality information — a channel quality indicator (CQI). Based on this CQI, the base station can adapt data rate based on channel conditions: the better the channel condition, the higher the data rate. Fast retransmission mechanism enables a mobile device to NACK each erroneous downlink packet in order to request a retransmission from its base station instead of the originating server. Opportunistic scheduling Most 3G data services implement an opportunistic scheduler (Both HSDPA [54] and EVDO [55] outline the use of an opportunistic scheduler in the downlink. Several service providers also confirmed the use of opportunistic scheduler in their data networks) . In cellular networks, channel conditions of mobile devices are time-varying and location-dependent. Since instantaneous channel conditions derive the instantaneous data rates of mobile devices [56], mobile devices periodically measure and report their CQIs to their base stations. An opportunistic scheduler at a base station selects a user with relatively good channel condition to transmit while maintaining predefined QoS or fairness constraints. Thus, opportunistic schedulers often achieve higher network performance than schedulers that do not take into account instantaneous channel conditions such as round robin. A very popular opportunistic scheduler is Proportional Fair (PF), whose design goal is to maximize the product of the throughput delivered to all users [52, 53]. In PF, each mobile device measures its instantaneous channel conditions through pilot signals, estimates the achievable data rate under its channel condition (denoted as CQIi (t) for user i at time t), and sends the information back to the base station. To achieve the goal of maximizing the product of the throughput delivered to all users [57], the PF scheduler chooses the user with the highest ratio of CQIi (t)/Ri (t) 2 , where Ri (t) is the average throughput of user i at time t. It is 2 PF

makes scheduling decisions based on the ratio DRCi (t)/Ri (t) where DRCi (t) = min{CQIk [n],

Bk [n] tT T I }

and Bk [n] is

CHAPTER 4. SCHEDULER ATTACK estimated by the base station as follows:   αCQI (t) + (1 − α)R (t − 1) if the user i is scheduled at t i i Ri (t) =  (1 − α)Ri (t − 1) otherwise

29

(4.1)

where α is a network provider’s parameter describing the weight of the current time slot toward the average. Typically, α is set as 0.001. Handoff Cellular networks implement handoffs to transfer a connection from one base station to another. There are two types of handoffs: soft and hard. In hard handoff, the network drops the connection to the current base station before initiating a new one. In soft handoff, on the other hand, a mobile device can have connections from several base stations simultaneously and choose to transmit through the best base station. Noticeably, handoffs in 3G cellular services do not break data transmission sessions.

4.1.2 Overview of attacks 3G data networks include mobile devices in their TCB. However, attackers can modify mobile devices to perform actions different from intended by the providers, even when providers attempt tamper-proof techniques [32, 58]. By trusting all mobile devices, 3G data networks suffer from at least two vulnerabilities. Fabricated CQIs Opportunistic schedulers base their scheduling decisions on CQIs reported by mobile devices without verification. By reporting fabricated CQIs, malicious mobile devices can manipulate the schedulers to achieve unfair network utilization and to disrupt other mobile devices. For instance, a malicious mobile device can report an inflated CQI such that its ratio of CQI to average data rate is the highest among all the devices in its cell, therefore ensuring that it will be scheduled in the next time slot. By repeating this strategy, the malicious device may obtain a large portion of slots in a short period of time, which causes large delay and delay jitter to other users (Section 4.2.1). Greedy handoffs Mobile devices may initiate soft handoffs, but opportunistic schedulers are oblivious of handoffs. For example, when a mobile device performs a handoff to another base the buffer size. We eliminated buffer dependence from our calculations for simplicity.

CHAPTER 4. SCHEDULER ATTACK

30

station, the new base station does not retrieve the device’s average data rate from its previous base station [40], but rather assigns an often small or average value as the device’s initial average rate. In the previous attack via reporting fabricated CQIs, the malicious mobile device has to report monotonically increasing CQIs to sustain the attack because its average data rate keeps increasing. Eventually, the attack becomes ineffective when its reported CQI exceeds the maximum allowable CQI. However, if the malicious device sits in the coverage of multiple base stations, it may handoff to another cell to acquire a fresh, lower average data rate and to start the attack again. Moreover, multiple malicious devices may cooperate to attack multiple cells simultaneously (Section 4.2.2).

4.2 Attack analysis Threat model Our threat model assumes that (1) attackers control one or a few mobile devices that a cellular network has admitted; and (2) attackers have modified the devices to report any CQI value to the base station and to initiate handoff at any time. We believe this threat model is realistic. Attackers can buy network-approved mobile devices and prepaid data plans or can spread worms to take over existing mobile devices. Moreover, experiences show that attackers can modify mobile devices to perform different actions than intended by the providers, even when providers attempt temper-proof techniques [32, 58, 59]. However, our threat model does not assume that attackers attack the cellular network infrastructure directly, e.g. by hacking into the network. Instead, they exploit vulnerabilities in the network’s scheduler by manipulating the information that their mobile devices report to the network. Attack settings

From this point on, we use attacker to refer to either a human adversary or the

mobile device of the adversary (the context should differentiate the two meanings), and use user to refer to either a human user or the mobile device of the user. When an attack involves multiple attackers, we assume that they coordinate. We will consider attacks on the proportional fair (PF) scheduler under three settings. First, we consider attacks from a single cell, with a single or multiple attackers. Next, we consider attacks from multiple cells, which is much more effective. Finally, we consider a more realistic situation where the attackers do not know the channel conditions of other users.

CHAPTER 4. SCHEDULER ATTACK

31

4.2.1 Attack within a single cell We consider the situation when all the attackers stay in the same cell. Starting with one attacker, we use mathematical analysis and simulation to evaluate his attack strategy. Then, we extend our analysis to multiple attackers in the same cell. We assume that no user leaves or joins the cell during the attack. Although this assumption is not crucial to our attack, it simplifies our analysis. We also assume that the attackers know the channel conditions of all the users in the cell. Section 4.2.3 will describe an attack strategy for the situations when this assumption does not hold. Single attacker The goal of the single attacker is to obtain a large number of consecutive time slots, therefore causing severe delay and jitter for the other victim users in the same cell. Since the PF scheduler assigns the next time slot to the user that has the highest ratio of instantaneous achievable data rate (measured in CQI) to average throughput, the attacker can report a large enough CQI to obtain the time slot. To obtain consecutive time slots, the attacker must report monotonically increasing CQIs (because its average throughput is increasing while other users’ throughput is decreasing, according to Equation 4.1) until its reported CQI exceeds the range of CQI values. It is difficult to calculate the precise number of consecutive time slots that the attacker can get, because the number depends on the channel conditions of all the users in the cell. However, we can estimate an upper bound of this number by considering a simplified situation where each user has the same CQI.3 First, we calculate the average throughput of a user. Let Ri (t) be the average throughput of user i at time slot t. Recall from Section 4.1.1 that   αCQI (t) + (1 − α)R (t − 1) if the user i is scheduled at t i i Ri (t) =  (1 − α)Ri (t − 1) otherwise

(4.2)

Since we assume that each user has the same CQI, the PF scheduler becomes a round robin scheduler, where each user is scheduled once every N slots (N is the number of users in the cell). For

example, if user i is scheduled at time slot s, he will not be scheduled until time slot s + N. Therefore, user i’s average rate Ri (t) maximizes at time slot s, and minimizes at the time slot s + N − 1. According to Equation 4.1, Ri (s) = (1 − α)N Ri (s − N) + αCQI 3 And

each user always has outstanding data to receive.

(4.3)

CHAPTER 4. SCHEDULER ATTACK

32

Let us consider a steady state, where Ri (t) = Ri (t + kN) for all integer k. In this case, Ri (s) = Ri (s − N). Using this equality in Equation 4.3, we have Ri (s) =

CQI αCQI ≈ 1 − (1 − α)N N

(4.4)

Ri (s) is user i’s maximum throughput. His minimum throughput is Ri (s − 1) = Ri (s + N − 1) = (1 − α)N−1 ∗ Ri (s) ≈ (1 − α)N−1

CQI N

(4.5)

Let C(t) = maxi {CQI/Ri (t)} be the maximum of CQI-to-throughput ratio at time t among all the users. In the steady state, C(t) becomes a constant C, which is: C=

CQI N ≈ Ri (s − 1) (1 − α)N−1

(4.6)

Next, we describe a strategy for the attacker to obtain consecutive time slots. To obtain time slot 1, the attacker i must report a CQIi (1) such that CQIi (1)/Ri (0) ≥ C(0). After time slot 1, C(1) = C(0)/(1 − α), because for each victim user j, its CQI remains constant, but its average throughput R j has been scaled by 1 − α. Therefore, to obtain time slot 2, the attacker i must report CQIi (2) such that CQIi (2)/Ri (1) ≥ C(1) = C(0)/(1 − α). Subsequently, at time t, the attacker must claim CQIi (t) such that CQIi (t)/Ri (t − 1) ≥ C(0)/(1 − α)t−1 . The attacker can obtain consecutive time slots until the required CQIi (t) exceeds CQImax , the maximum value of CQI. Therefore, the maximum number of consecutive time slots that the attacker can obtain is the maximum integer t0 that satisfies  t0 −1  C αC CQImax ≥ Ra (0) ∏ + (1 − α) k−1 (1 − α)t0 −1 k=1 (1 − α)

(4.7)

Equation (4.7) shows that the maximum number of consecutive slots an attacker can obtain (t0 ) depends on the average throughput of the attacker at the beginning of the attack (Ri (0)), the maximum CQI (CQImax ), and α. Since the maximum CQI and α are set by the system, they are out of the control of the attacker. The maximum CQI depends on the hardware. α is used to balance the tradeoff between long-term and short-term performance. The smaller the value α, the better the system’s long-term throughput; however, when under attack, the smaller the value α, the larger the value of t0 , i.e., the attacker can obtain more time slots. By comparison, the attacker has control over Ri (0), its average throughput at the beginning of the attack. Equation (4.7) shows that the smaller the value Ra (0), the larger the value t0 . Therefore, after each attack session, the attacker needs to reset its Ra (0) by reporting lowest CQI values for a sufficient period (typically on the order

CHAPTER 4. SCHEDULER ATTACK

33

of seconds). Finally, this model is simplified, assuming all victim users have the same, consistent CQI. When users have users have time-varying channel conditions, Equation 4.7 provides an upper bound for estimating t0 . Multiple attackers A single attacker can obtain consecutive time slots until his reported CQI exceeds the maximum CQI value; however, we can increase the number of consecutive time slots obtained by using multiple colluding attackers. We describe three different coordinating schemes. Sequential attack

The simplest scheme is to attack sequentially. The attacker with the smallest

average throughput Ri (t) starts the attack and tries to obtain as many consecutive time slots as possible, while the other attackers lurk (by reporting arbitrarily small CQIs to avoid being scheduled). When the active attacker’s reported CQI exceeds the maximum value of CQI, it stops the attack while the attacker with the smallest average throughput starts to attack. The attack continues until no attacker can get scheduled (because their average throughput is too high). Minimum CQI Attack

Since the attack will stop when all attackers’ reported CQIs exceed the

maximum value, this scheme tries to slow the increment of the reported CQIs. At each time slot, each attacker computes the CQI that it needs obtain the time slot. Then, the attacker with the smallest CQI reports its CQI to the base station while the other attackers lurk. Delta CQI Attack

This algorithm tries to slow the increment of calculated CQI values. At each

time slot t, each attacker i computes the increment δi (t) needed to its previous CQI. In other words, δi (t) = CQIi (t) − CQIi (t − 1). The attacker with the smallest δi (t) then reports its CQI to the base station. Simulation We used simulation to evaluate the effectiveness of our attacks in a single cell. In the simulation, we chose parameters that were recommended by specifications or that were commonly used by cellular networks. The PF scheduler had α = .001. The cell had 50 users. Each user quantized his channel condition into CQI, an integer between 1 and 15, and reported the CQI to the base station. The goal of the attack was to obtain the maximum number of consecutive time slots.

CHAPTER 4. SCHEDULER ATTACK

34

First, we simulated a single attacker in a cell with 49 victim users. We used the same ideal scenario as in our analysis in Section 4.2.1, i.e., all victim users had the same CQI value. The simulation showed that the attacker could obtain 42 consecutive time slots, whereas Equation 4.7 predicts that the attacker can obtain 39 consecutive time slots. The minor difference between the simulation and the analysis is due to the approximation during the derivation of Equation 4.7. Next, we simulated the same attack under a more realistic condition where each user’s channel condition was a random variable following a Rayleigh distribution with σ = 3 and an initial average rate of 0.5. The simulation showed that the attacker gained an average of 19 time slots, with a standard deviation of 2.77. Next, we simulated multiple attackers in the same cell. Again, each user’s channel condition was a random variable following a Rayleigh distribution. We varied the number of attackers from one to five and simulated each of the attack schemes in Section 4.2.1. Figure 4.1 shows that the number of collective consecutive time slots obtained by the attackers increases almost linearly with the number of attackers. Among the three attack schemes, the Delta CQI scheme performed the best, where five attackers obtained 99 consecutive time slots.

Timeslot Occupied

100

80

60

40

20

0 0

Delta Minimum Sequential 1

2

3

4

5

Number of Attackers Figure 4.1: Consecutive time slots obtained By attackers using different collaborating schemes in a single cell.

CHAPTER 4. SCHEDULER ATTACK

35

Although 99 consecutive time slots (or 165ms) occupied by the attackers will cause delay on victim users, this delay is tolerable by many applications and protocols. Moreover, after the attack, the attackers must relinquish a large number of (at least 2000) time slots to reset their average throughput low enough before they can attack again. Therefore, the attackers cannot sustain this delay. Fortunately (or unfortunately, depending on your stand), we were able to exploit another vulnerability to make our attack much more effective and sustainable.

4.2.2 Attack from two cells When attackers sit in the overlapping area of two cells, they can exploit handoff to make their attack much more effective and sustainable. Section 4.2.1 shows that an attacker’s reported CQI and average throughput increase very fast during an attack. When a large average throughput forces the attacker to report a CQI larger than its maximum value, the attack stops. However, since users can initiate soft handoff and the network does not carry users’ average throughput across cells, the attacker can hand off to another other cell, get a small initial average throughput, and immediately start the attack in its new cell. Initial average throughput Since the network does not track users’ average throughput across cells [40], when a new user joins a cell, the scheduler must first assign the user an initial value for its average throughput. Since the choice of this initial value is unspecified, we explore three reasonable schemes. Although these schemes are not all-inclusive, they represent good schemes that lead to predictable behavior of the PF scheduler. Based on the average of average throughput of all users

A simple scheme is to choose the

average of average throughput of all existing users in this cell as the initial average throughput of the new user. The motivation for this scheme is the assumption that the new user’s channel condition is close to the average channel condition of all existing users. The disadvantage of this scheme is that when the new user just moves into his current cell from a neighboring cell, he is likely at the edge of his current cell with poor channel condition, so this scheme would over-estimate the new user’s average throughput.

CHAPTER 4. SCHEDULER ATTACK Based on the minimum of average throughput of all users

36 Since new users often join a cell

from the edge of the cell, they expect to have the poorest channel condition. Therefore, this scheme chooses the minimum of the average throughput of all existing users as the initial average throughput of the new user. However, if the new user happens to have good channel condition, this scheme would under-estimate the user’s average throughput. Determined by the user Finally, since users are burdened with tasks such as channel quality and pilot measurements for multiple cells, an intuitive scheme is to let users report their initial average throughput. A major problem with this scheme is that the base station trusts users blindly. An attacker can report a bogus low average throughput to gain unfair advantage in scheduling. Simulations We simulated the attack from two cells. We used the same PF scheduler and the same Raleigh distribution for users’ channel conditions as in Section 4.2.1. We simulated various number of attackers per cell, from one to five. The attackers used the sequential attack algorithm described in Section 4.2.1. However, after the last attacker in the cell finished his attack (i.e., when he could obtain no more time slots), all the attackers in both cells handed off to the other cell. Although the sequential attack algorithm is not the most effective, it is the simplest and illustrates the lower bound of the attack effect. We assume that handoff takes one time slot, which is realistic for soft handoff. We ran the simulation for 18072 time slots, or 30 seconds. Figure 4.2 shows the percentage of time slots that the attackers got where there was one attacker per cell and the attackers determined their initial throughput. It shows that after about 2000 time slots, the attackers consistently obtained about 78% of all the time slots, a condition that we call the stabilization of the attack. We simulated different number of attackers per cell and different schemes for assigning the initial average throughput, and in all the simulations the attack stabilized well before 30 seconds. Figure 4.3 shows the total number of time slots that the attackers obtained in 30 seconds. Unsurprisingly, the more attackers per cell, the more time slots that they obtained. However, even with just one attacker per cell, the attackers obtained from 13459 (74%) to 16241 (90%) time slots, depending on the scheme by which the scheduler assigns the initial average throughput. Among the three schemes, the scheme that let the user provide this initial value is the most vulnerable, where one attacker obtained 16241 (90%) time slots while five attackers obtained 17317 (96%) time slots.

CHAPTER 4. SCHEDULER ATTACK

37

% of Timeslots obtained

1

0.9

0.8

0.7

0.6

0.5 0

5000

10000

15000

Time (timeslot) Figure 4.2: Percentage of time slots obtained by two attackers, one per cell

4.2.3 Attack without knowing victims’ CQIs So far, our attack requires the attackers to know all users’ channel conditions and average throughput at each time slot. In practice, however, attackers may not have such information. In this case, the attacker must predict the maximum of the CQI-to-throughput ratios of all the victim users, because the attacker can obtain the next time slot only if his own ratio is larger than those of all the victim users. We show how the attacker can predict the value. First, the attacker performs a statistical analysis of the PF scheduler and estimates a distribution of the max ratio of all users, shown in Figure 4.4. Based on this distribution, the attacker determines the CQI that he needs to obtain the next time slot with high probability. The higher CQI that the attacker reports, the higher probability that he will obtain the next time slot; however, in this case, he will also exceed the maximum CQI value more rapidly, which will force him to hand off to another cell. Since the attacker cannot obtain time slots during handoff, he must strike a balance between obtaining the next time slot with high probability (by reporting higher CQIs) and reducing the frequency of handoffs (by reporting lower CQIs). In the simulation, we choose the 90% value of the c(t) in Figure 4.4. In other words, the attacker has 90% chance of obtaining a time slot. During the attack, the attacker must adjust the estimated maximum CQI-to-throughput

CHAPTER 4. SCHEDULER ATTACK

38

18000

Timeslots Occupied

16000 14000 12000 10000 8000 6000 4000

User Provided Min. user avg. Mean user avg.

2000 0 0

1

2

3

4

5

Attackers per Cell Figure 4.3: Time slots occupied by the attack in 30 seconds (18072 time slots). The attackers have full knowledge of every user’s information in each cell. The three lines represent different schemes for assigning the initial average throughput by the base station ratio of all the victim users constantly. This is because that in every time slot, each user’s average throughput will increase by α∗CQI if he is scheduled, and decrease by a factor of (1− α) otherwise. We propose the following scheme for adjusting the maximum ratio estimation. Let c(t) be the estimated maximum CQI-to-throughput ratio at time t, derived from the distribution in Figure 4.4, and Ri (t) be the average throughput of user i at time t. If the attacker is scheduled in time t, the average throughput of all the other users will decrease, Ri (t) = (1 − α) ∗ Ri (t − 1). Since c(t) estimates the largest Ri (t) of all the victim users, it increases at the same rate, c(t + 1) = c(t)/(1 − α). When the attacker is not scheduled, on the other hand, only the average rate of the victim user who is scheduled will increase. Therefore, CQIi (t + 1) CQIi (t + 1) ≈ max i Ri (t − 1) ∗ (1 − α) + α/N ∗CQIi (t) Ri (t) CQIi (t + 1)/Ri (t − 1) c(t) = max ≈ i (1 − α) + α/N ∗CQIi (t)/Ri (t − 1) (1 − α) + α/N ∗ c(t)

c(t + 1) = max i

Some approximations are involved in the above estimation. First, on average, a victim user gets scheduled once every N times when the attacker is not scheduled. Therefore, the average rate of a victim user will be increased by α/N ∗CQIi (t) when the attacker is not scheduled approximately.

CHAPTER 4. SCHEDULER ATTACK

39

Number of times scheduled

700 600 500 400 300 200 100 0 30

40

50

60

70

80

90

Scheduled CQI/R ratio Figure 4.4: Histogram of the CQI-to-throughput ratio of the scheduled user over 10000 time slots Second, when a user is scheduled, his CQI-to-throughput ratio is the maximum among all users and thus its value of CQIi (t)/Ri (t − 1) is approximately c(t). Equation 4.8 summarizes our analysis:   c(t)/(1 − ε) if the attacker is scheduled c(t + 1) = (4.8)  c(t)/(1 + σ ∗ (c(t) − 1)) if the attacker is not scheduled where ε and σ are functions of α. We used ε and σ instead of α to compensate for the possible errors in our estimation of the maximum CQI-to-throughput ratio, and we determined them empirically. We reran the simulations in Section 4.2.2 but with the estimated maximum CQI-to-throughput ratio using Equation 4.8. As we experimented with different choices of ε and σ, we noticed that over-estimation of the maximum CQI-to-throughput ratio yielded better attack results than underestimation. Figures 4.5(a) shows the number of time slots obtained by the attackers with ε = 1.5 ∗ α and σ = α/60. When there is a single attacker per cell, the attackers may obtain between 11583 (64%) and 15874 (88%) time slots, depending on the scheme for assigning initial average throughput. When there are five attackers per cell, they can obtain between 14353 (79%) and 17136 (95%) time slots. Next, we compare the effect of the attack under this realistic situation where attacks do not know the CQIs of the victims with the effect of the attack under the ideal situation. Figure 4.5(b)

CHAPTER 4. SCHEDULER ATTACK

40

shows the percentage of time slots that the attackers obtain when they do not know the CQIs of the victim users, compared to the case with perfect information. If the PF scheduler uses user-provided initial average throughput, the attackers, using our estimation, can obtain almost the same number of time slots, as in the ideal situation when the attackers know the value of c(t) perfectly. Even when the PF scheduler uses the two other schemes, which are more robust against our attack, the attackers could still obtain more than 85% of the time slots that they would obtain in the ideal situation.

4.3 Attack impact In this section, we discuss the impact of our attacks on normal victim mobile users. We first produce a benchmark from an honest user. Then, we recount a study on a particular application, VoIP, and show how the attack can render VoIP useless. Overall and individual average throughput Figure 4.6 compares the distribution of users’ throughput in a cell before and during the attack described in Section 4.2.2. We calculate the throughput based on HSDPA’s mapping of CQI to data rate [60, 61]. Before the attack, most users obtain an average throughput of 40 to 55 kbps. During the attack, most users obtain an average throughput of only 10 to 15 kbps. By comparison, the attacker obtains an unusually high average throughput of 1.5Mbps4 5 . This figure shows that the attack has defeated the fairness objective of PF entirely. Average user delay

Figure 4.7 depicts the average delay between each transmission for victim

users. In a normal application, users experience a slight delay of 0.081 seconds between each transmission, which is acceptable in most applications. This number, by the way, would most likely be different if QoS requirements are deployed. Note that PF scheduler functions very similar to a round robin scheduler in this case, even with a random channel condition. For example, in a cell with 50 users, each user waits around 49 timeslots for a transmission. The delay variance of a user in a PF scheduler is higher (in particular, N 2 − N where N is the number of users). During the attack, on the other hand, a victim user can experience up to 1.8 seconds delay (22 times the delay before the attack) between each pair of successive transmissions. This significant delay will render many applications virtually useless. 4 Not

shown in graph due to space constraints reason why the attacker’s average throughput is not significantly higher is because of handoffs, which resets the attacker’s average throughput. 5 The

CHAPTER 4. SCHEDULER ATTACK Impacts on applications

41

Cellular providers have already started offering the VoIP service due

to its lower cost. However, VoIP packets have a rigorous delay requirement: 0-150ms delay is acceptable, 150ms-400ms delay might be tolerable, but longer delay is disruptive [62]. This delay budget is for end-to-end delay, including coding/decoding time (around 20ms), transmission delay over the Internet (about 100ms across the continental USA) and uplink and downlink delay to the user. Therefore, the delay on the cellular link is important. Using VoIP as an example, we develop a function to calculate the end-to-end delay between two cellular users. Let f (DNW, N,U, X ) be the end-to-end, one way delay between two users, where DNW is the backbone network delay, N is the number of users in a cell, U is the uplink network delay, and X is the delay that the attack induces to the downlink. Equation 4.9 shows the formula for EV-DO, based on Goode’s calculation [63].

f (DNW, N,U, X ) = 101.1ms + DNW ms + (N)(1.67ms) +U ms + X ms

(4.9)

Recall that we have shown in Figure 4.7 that honest users can experience up to 1.8 seconds of delay between each transmission. Using equation 4.9, we obtain 2094.6ms of end-to-end delay per packet for every honest user in the cell. This delay is devastating for VoIP communication as it renders it useless. Furthermore, the attack can be extented indefinitely with only 5 attackers comprising only 10% of total number of users in the cell.

4.4 Possible defense strategies The above-illuminated attack exploits several vulnerabilities that in combination, enable any user to perform a denial-of-service attack on downlink cellular data service. To defend against this type of attacks, we must either eliminate these vulnerabilities or mediate them such that this type of attack is no longer effective. Here we outline some defense strategies that could be implemented either in future specification revisions or current implementations.

4.4.1 Attack detection There are two characteristics that the base station can use to distinguish normal and underattack operations in the attack outlined in section 4.2; namely, decrease of average user throughput and excessive numbers of handoffs. We can take advantage of these characteristics in composing a defense strategy.

CHAPTER 4. SCHEDULER ATTACK

42

Anomaly detection using average throughput in a normal condition

The base station can mea-

sure an average user’s throughput during normal operation, either by simulation or by actual measurement. Then, it can compare the current throughput with recorded normal throughput. If their difference is above a certain threshold, this could indicate that the system is under attack. The base station then can use several methods to mitigate the attack, including using a scheduler that does not require user collaboration, such as round robin, and tracing the attack source. Number of handoffs per user In a normal operation, users do not perform excessive handoffs. On the other hand, attackers performs handoffs as many as one every 5 time slots. The base station can observe the number of handoffs performed per user over a period of time. If a user performs an unusually high number of handoffs in a given time, the base station can reject further handoff requests, thereby stopping the attack in that cell.

4.4.2 Attack prevention In addition to attack detection, network designers can implement modifications to the existing architecture to prevent the attack. We present some possible mitigation strategies that are designed to stop this attack from causing significant damage. Trust but verify

The root of this attack is that the cellular network considers mobile users in

its TCB, and subsequently makes decisions based on unverified user input. One possible defense strategy that the base station can perform is to periodically check the validity of the various reports made by the mobile devices. This approach, trust, but verify, places a level of trust on a mobile device as it passes random checks. The CQI reported by the mobile device can be randomly checked using their uplink channel condition and error rate. 6 . Mobile station gains trust as it passes random check, and loses trust when it fails. The base station can disconnect mobile users if they fail checks for a number of times thereby punishing malicious or badly configured users. Assume an average based on the normal operating condition

The base station can define an

average such that it reflects the normal behavior of the cell, given the number of users in the cell. This scheme punishes the attacker because he will be assigned a high average each time he switches cells and thus force the attacker to perform handoffs much earlier. Figure 4.8 illustrates that this 6 Estimating

claims

channel condition using uplink data rate is not perfectly accurate; however, it can still detect anomalous

CHAPTER 4. SCHEDULER ATTACK

43

scheme is much more resilient to attack than the strategy that solely relies on the cell’s current information. This scheme does not require a change to the current architecture nor does it require collaboration between network entities. This scheme is not without cost. During handoff, an honest user is usually at the edge of the cell with relatively poor channel condition. Assigning a fixed average may starve the user for a while. Additionally, while this scheme alleviates the effect of attacks by forcing more handoffs, it does not solve the whole problem. Scheduler information sharing

The attack discussed in Section 4.2.2 accentuates the fact that

PF is oblivious of handoffs. When a mobile device performs a handoff the base station must derive the mobile users’ average throughput using information confined to the current cell. Therefore, whenever the attacker performs a handoff, the attacker’s excessively high average throughput gets replaced with a new, calculated average throughput from the new base station. The attack can be stopped if base stations communicate and transfer users’ average throughput along with a mobile device during handoff. In this case, the attacker cannot continue the attack for more than a few slots since the average throughput is extremely high. This defense mechanism can limit the effect of the attack to that in a single cell and therefore reduce the problem to service degradation instead of denial of service. Figure 4.8 shows the performance of this defense strategy compared to the average throughput assignment strategy, which does not require collaboration. Notice that by sharing information between cells, the attack can be significantly deterred. Modified schedulers with multi-dimensional constraints

Another possible defense strategy is

to introduce additional constraints into the scheduler. These temporal constraints limits the portion of time slots allocated to the attackers, stopping the attack. The base station can implement two types of temporal constraints, either individually, or combined. The long-term temporal constraint limits the minimum and (possibly) maximum portion of time slots allocated to each user over a long period of time (usually during the lifetime of a user, on the order of minutes). The short-term temporal constraint guarantees that each user obtains a minimum (and possibly maximum) number of time slots within a time window, on the order of a few hundred milliseconds. Long-term temporal constraints can be easily satisfied due to the PF scheduler’s fairness constraints in a normal operation. The impact of the short-term constraint, on the other hand, depends on the parameters used. The defense capability becomes more effective as the number of slots

CHAPTER 4. SCHEDULER ATTACK

44

assigned to each user in a short-term increases, but lowers overall throughput 7 . However, shortterm constraint is also useful in improving short-term fairness among users and reduce throughput burstiness, which is desirable to upper layer protocols, such as TCP. Priority queue Additionally, a priority queue can be implemented at the base station. Traffic with delay constraints, such as VoIP traffic, can be scheduled with high priority, while other traffic, such as web browsing, can be scheduled with low priority. While an attacker can claim to be high priority, because the number of high priority users is relatively small, these users have much better delay performance, and thus mitigate the effects of the attacks (in particular, attacks without handoff). In addition, the priority scheme may be combined with temporal constraints so that an attacker cannot claim a large portion of resource (for attacks with handoff).

4.5 Conclusion In this chapter, we have shown that cellular data networks are vulnerable to DoS attacks because of the following vulnerabilities: • The network trusts mobile devices to report truthful CQIs, which the PF scheduler uses without verification for assigning time slots. Therefore, malicious mobile devices can manipulate their reported CQIs to gain a large number of time slots. • The network does not track the average throughput of mobile devices across different cells, which allows malicious devices to maintain perpetual scheduling priority by frequent handoffs. We have studied a series of attacks on the proportional fair scheduler exploiting the above vulnerabilities. Our simulations show that just one attacker per cell can disrupt time-sensitive data services, such as voice-over-IP. Moreover, multiple attackers in the same cell can collaborate to cause serious denial of service by occupying up to 95% of scheduling slots indefinitely. Meanwhile, they can also induce 1.8s delay between each consecutive packet transmission on every victim user who is in the same cell as the attackers. We have proposed several mitigation strategies to defend against these attacks. 7 because

the scheduler has to schedule users given the short-term constraint although there might be a user with a higher value of CQI/R

CHAPTER 4. SCHEDULER ATTACK

45

18000

Timeslots Cccupied

16000 14000 12000 10000 8000 6000 4000

User provided Min. user avg. Mean user avg.

2000 0 0

1

2

3

4

5

Attackers per Cell (a) Time slots obtained by the attack in 30 seconds (18072 slots). The attackers do not have any knowledge of victims’ CQIs.

Accuracy of Prediction (%)

100

80

60

40

20

0 0

User Provided Min. user Avg. Mean user Avg. 1

2

3

4

5

Attackers per Cell (b) Percentage of time slots obtained by attackers without knowing victims’ CQIs compared to those with knowing victims’ CQIs

Figure 4.5: Performance of the attack without knowing victims’ CQIs. Each sub-figure shows three curves, each representing a different scheme for assigning the initial average throughput.

CHAPTER 4. SCHEDULER ATTACK

46

25

% of Users

20

15

10

After Attack 5

0 0

10

20

30

40

50

60

Average Throughput (kbps) Figure 4.6: The average throughput distribution of users, before and after attack.

CHAPTER 4. SCHEDULER ATTACK

47

Victim’s Average Delay (s)

2.5 Full Knowledge Prediction 2

1.5

1

0.5

0 0

1

2

3

4

5

Attackers per Cell Figure 4.7: The average delay an honest user experiences between each packet transmission .

CHAPTER 4. SCHEDULER ATTACK

48

18000

Timeslots Occupied

16000 Min. User Avg. Mean Norm. User Avg. Collaboration

14000 12000 10000 8000 6000 4000 2000 0 0

1

2

3

4

5

Attackers per Cell Figure 4.8: Continuous number of timeslots occupied by the attack using different defense strategies compared to a normal attack in full knowledge

CHAPTER 5. SUMMARY AND CONCLUSION

49

Chapter 5

Summary and Conclusion While cellular users embrace new broadband data services and applications, attackers get opportunities to exploit emerging vulnerabilities in mobile devices, cellular data networks, and the interaction between cellular networks and the Internet. We demonstrated two particular denial of service attacks that can cause havoc in cellular network with relatively few resources. In the first attack in Chapter 3, we demonstrated drains mobile devices’ battery power as much as 22 times faster. This attack proceeds in two stages. First, the attack exploits vulnerabilities in MMS to build a hit list of mobile devices. Then, the attack exploits PDP content retention and the paging channel to drain mobile devices’ battery power. We were able to drain batteries without alerting either the mobile user victims or the cellular network operators. Our analysis shows that an attacker would need only several home DSL Internet connections to mount a large scale attack against a large number of cellular phones. We identified key components in cellular networks that enable this attack and proposed corresponding mitigating solutions. In the second attack in Chapter 4, we studied a series of attacks on the proportional fair scheduler exploiting the 3G and PF vulnerabilities. Our simulations show that just one attacker per cell can disrupt time-sensitive data services, such as voice-over-IP. Moreover, multiple attackers in the same cell can collaborate to cause serious denial of service by occupying up to 95% of scheduling slots indefinitely. Meanwhile, they can also induce 1.8s delay between each consecutive packet transmission on every victim user who is in the same cell as the attackers. We have proposed several mitigation strategies to defend against these attacks. Due to the complex interaction between mobile devices, cellular data networks, and the Internet, we conjecture that our discovered attacks may be just the tip of an iceberg. We hope that our work will bring attention to this emerging threat and will inspire future research for securing

CHAPTER 5. SUMMARY AND CONCLUSION

50

cellular data services and applications. Furthermore, we hope to motivate the cellular industry to improve the security in their current data services, and to scrutinize the security in their future specifications more rigorously.

BIBLIOGRAPHY

51

Bibliography [1] CITA.

Wireless quick facts.

http://files.ctia.org/pdf/Wireless Quick Facts

October 05.pdf. [2] Charles Brookson. GSM ( and PCN ) security and encryption. http://www.brookson.com/ gsm/gsmdoc.htm. [3] Paul Yousef. GSM-security: a survey and evaluation of the current situation. Master’s thesis, Linkoping Institute of Technology, 2004. [4] Chengyuan Peng. GSM and GPRS security. In HUT TML, 2000. [5] Alan Bavosa. GPRS security threats and solution recommendations. http://www.juniper. net/solutions/literature/white papers/200074.pdf. [6] Charles Brookson. GPRS security. http://www.brookson.com/gsm/gprs.pdf. [7] Oillie Whitehouse.

GPRS security:

Not ready for prime time.

http://www.

securitymanagement.com/library/wireless tech0902.pdf. [8] Stephane piot. Security over GPRS. Master’s thesis, University College London, 1998. [9] Christopher Wingert and Mullaguru Naidu. CDMA 1XRTT security overview. http://www. cdg.org/technology/cdma technology/white papers/cdma 1x security overview. pdf. [10] Atique Ahmed Khan. Security and vulnerability analysis of wireless messaging protocols and applications. In Pak Con, 2004. [11] Stefan Andersson. MMS security considerations. In 3GPP TSG SA WG3 Security, 2003.

BIBLIOGRAPHY

52

[12] Rei Safavi-Naini, Willy Susilo, and Gelareh Taban. Towards securing 3G mobile phones. In The 9th IEEE International Conference on Network (ICON 2001), 2001. [13] 3rd Generation Partnership Project. 3G security: Security threats and requirements. ftp: //ftp.3gpp.org/Specs/2000-12/R1999/21 series/21133-310.zip. [14] Ollie Whitehouse and Graham Murphy.

Attacks and counter measures in 2.5G and 3G

cellular IP networks. http://www.atstake.com/research/reports/acrobat/atstake cellular networks.pdf. [15] Sotirios I. Maniatis Constantinos F. Grecas and Iakovos S. Venieris. Introduction of the asymmetric cryptography in GSM, GPRS, UMTS, and its public key infrastructure integration. In Mobile Network and Applications, 2003. [16] Chi-Chun Lo and Yu-Jen Chen. A secure communication architecture for GSM networks. In IEEE Transactions on Consumer Electronics, 1999. [17] John A. MacDonald and Chris J. Mitchell. Using the GSM/UMTS SIM to secure web services. In the 2nd Workshop on Mobile Commerce and Services WMCS, 2005. [18] ISAAC. GSM cloning. http://www.isaac.cs.berkeley.edu/isaac/gsm.html. [19] M. J. Riezenman. Cellular security: better, but foes still lurk. IEEE Spectrum, 37(6), 2000. [20] B Sun, F Yu, K Wu, and VCM Leung. Mobility-based anomaly detection in cellular mobile networks. In 2004 ACM workshop on Wireless security, 2004. [21] Mirela Sechi Moretti Annoni Notare, Fernando Augusto da Silva Cruz, Bernardo Gonalves Riso, and Carlos Becker Westphall. Security management against cloned cellular telephones. In IEEE International Conference on Networks, page 356, Washington, DC, USA, 1999. IEEE Computer Society. [22] Azzedine Boukerche and Mirela Sechi M. Annoni Notare. Behavior-based intrusion detection in mobile phone systems. Parallel and Distributed Computing, 62(9):1476 – 1490, 2002. [23] Nilesh Agarwal, Leena Chandran-Wadia, and Varsha Apte. Capacity analysis of the GSM short message service. In National Conference on Communications, 2004.

BIBLIOGRAPHY

53

[24] William Enck, Patrick Traynor, Patrick McDaniel, and Thomas La Porta. Exploiting open functionality in SMS-capable cellular networks. In 12th ACM Conference on Computer and Communications Security (CCS’05), November 7-11, 2005. [25] Patrick Traynor, William Enck, Patrick McDaniel, and Thomas La Porta. Mitigating attacks on open functionality in SMS-capable cellular networkss. In 12th Annual International Conference on Mobile Computing and Networking MOBICOMM, 2006. [26] Pars Mutaf and Claude Castelluccia. Insecurity of the paging channel in the wireless internet: A denial-of-service attack that exploits dormant mobile IP hosts. In 3rd Workshop on Applications and Services in Wireless Networks, 2003. [27] Thomas Martin, Michael Hsiao, Dong Ha, and Jayan Krishnaswami. Denial-of-service attacks on battery-powered mobile computers. In Proceedings of the 2nd IEEE Pervasive Computing Conference, 2004. [28] Daniel C. Nash, Thomas Martin, Dong Ha, and Michael Hsiao. Towards an intrusion detection system for battery exhausion attacks on mobile computing devices. In Proceedings of 2nd International Workshop on Pervasive Computing and Communications Security (PerSec ’05), 2005. [29] Anti Phishing Working Group.

What is phishing and pharming?

http://www.

o2 germany promotes SMS-phishing.

http://www.

antiphishing.org/. [30] Redteam.

Advisory:

redteam-pentesting.de/advisories/rt-sa-2005-009.txt. [31] Brian Fonseca. Worm calling. http://www.infoworld.com/articles/hn/xml/00/06/ 06/000606hnphoneworm.html. [32] Dan Ilett and Matt Hines.

Skulls program carries cabir worm into phones.

http://news.com.com/Skulls+program+carries+Cabir+worm+into+phones/ 2100-7349 3-5469691.html. [33] James W. Mickens and Brian D. Noble. Modeling epidemic spreading in mobile environments. In WiSe ’05: Proceedings of the 4th ACM workshop on Wireless security, pages 77–86, New York, NY, USA, 2005. ACM Press.

BIBLIOGRAPHY

54

[34] R. Knopp and P. Humblet. Information capacity and power control in single-cell multiuser communications. In Proceedings of the ICC, 1995. [35] David Tse and Pramod Viswanath. Fundamentals of Wireless Communication. Cambridge, 1 edition, 2005. [36] X. Liu, E. K. P. Chong, and N. B. Shroff. A framework for opportunistic scheduling in wireless networks. Computer Networks, 41(4):451–474, March 2003. [37] Mohamad Assaad, Badii Jouaber, and Djamal Zeghlache. Effect of TCP on UMTS-HSDPA system performance and capacity. 2004. [38] Jin-Hee Choi, Jin-Ghoo Choi, and Chuck Yoo. Analyzing the impact of proportional fair scheduler on TCP performance. 2005. [39] Matthew Andrews. Instability of the proportional fair scheduling algorithm. In IEEE Transactions on Wireless Communications, 2004. [40] Tian Bu, Li Li, and Ramachandran Ramjee. Generalized proportional fair scheduling in third generation wireless data networks. In INFOCOMM, 2006. [41] Kameswari Kotapati, Peng Liu, Yan Sun, and Thomas F. La Porta. A taxonomy of cyber attacks on 3G networks. In Technical Report NAS-TR-0021-2005, Network and Security Research Center, Department of Computer Science and Engineering, Penn State University, 2005. [42] Fabio Ricciato. Unwanted traffic in 3G networks. In ACM SIGCOMM Computer Communication Review, Volume 36, Issue 2, 2006. [43] A Bovosa. Attacks and counter measures in 2.5G and 3G cellular IP networks. In Juniper White Paper, 2004. [44] Ashwin Sridharan, Ramesh Subbaraman, and Roch Guerin. Uplink scheduling in the EV-DO rev. a system: An initial investigation. In Sprint ATL Research Report Nr. RR06-ATL-080139, 2006. [45] C. Rose C.U. Saraydar. Minimizing the paging channel bandwidth for cellular traffic. In IEEE ICUPC, 1996.

BIBLIOGRAPHY

55

[46] Peter McGuiggan. GPRS In Practice: A companion to the specification. John Willey & Sons, 2004. [47] W3C Proposed Recommendation. Synchronized multimedia integration language (SMIL2.1). http://www.w3.org/TR/SMIL2/. [48] Humpa. MMS pic server. http://www.humpa.com. [49] Curt Schurgers Vijay Raghunathan, Saurabh Ganeriwal and Mani Srivastava. WFQ: An energy efficient fair scheduling policy for wireless systems. In ISLPED, 2002. [50] D. Moore, C Shannon, G. Voelker, and S. Savage. Inferring internet denial of service activity. In USENIX Security Symposium, 2001. [51] ETSI. GSM 03.60 general packet radio service: Service description, stage 2. http://webapp. etsi.org/workprogram/Report WorkItem.asp?WKI ID=3068. [52] A. Jalali, R. Padovani, and R. Pankaj. Data throughput of CDMA-HDR a high efficiencyhigh data rate personal communication wireless system. In Proceedings of IEEE Vehicular Technology Conference 2000-Spring, volume 3, 2000. [53] E. F. Chaponniere, P. Black, J. M. Holtzman, and D. Tse. Transmitter directed multiple receiver system using path diversity to equitably maximize throughput. U.S. Patent No. 6449490, 2002. [54] Harri Holma and Antti Toskala. HSDPA/HSUPA for UMTS. John Willey & Sons, 2006. [55] Vieri Vanghi, Aleksandar Damnjanovic, and Branimir Vojcic. The cdma2000 System for Mobile Communications. Prentice Hall, 2004. [56] S. Nanda, K. Balachandran, and S. Kumar. Adaptation techniques in wireless packet data services. IEEE Communications Magazine, 38(1):54–64, January 2000. [57] F. Kelly. Charging and rate control for elastic traffic. European Transactions on Telecommunications, 8:33–37, 1997. [58] Telefono. Homebrew mobile phone club. http://telefono.revejo.org/. [59] Radmilo Racic, Denys Ma, and Hao Chen. Exploiting MMS vulnerabilities to stealthily exhaust mobile phones’ battery. In IEEE SecureComm, 2006.

BIBLIOGRAPHY

56

[60] 3GPP. UMTS MAC protocol specification specification 3GPP TS 25.321 version 7.00 release 7. http://3gpp.org/ftp/Specs/html-info/25321.htm. [61] Troels E. Kolding, Frank Frederiksen, and Preben E. Morgensen. Performance aspects of WCDMA systems with high speed downlink packet access (HSPDA). In Vehicular Technology Conference, 2002. [62] ITU-T. One-way transmission time. ITU-T Recommendation G.114, 1996. [63] Bur Goode. Voice over internet protocol (VoIP). In IEEE, 2002.