Practical Traffic Analysis Attacks on Secure Messaging Applications - PowerPoint PPT Presentation

Electrical and Computer Engineering Department Practical Traffic Analysis Attacks on Secure Messaging Applications Alireza Bahramali, Ramin Soltani, Amir Houmansadr, Dennis Goeckel, Don Towsley University of Massachusetts Amherst

Instant Messaging is Popular! ❖ Over 2 billion people use Instant Messaging (IM) applications ❖ Used to exchange various types of messages 2

Typical IM Providers ❖ A variety of IM services: Telegram, WhatsApp, Signal ❖ Most IMs have centralized structure ➢ All the communications are relayed through IM provider servers IM Server 3

Typical IM Providers ❖ Various types of communication: ➢ One-to-one communication ➢ Group communication ➢ Channel communication: admins and members IM Server Admins Members 4

IM Communications are Sensitive ❖ Extensively used to exchange politically and socially sensitive contents ❖ Therefore, IM services are attractive targets for government and corporation surveillance 5

Examples 6

How Confidential Are IMs? The good news: content is protected by Encryption, End-to-Middle or End-to-End IM Server Client Client The bad news: traffic patterns leak information 7

How Patterns Leak? Admin Member1 Member 2 Member 3 8

Objective of this study: investigate the threat of traffic analysis to popular IM services ❖ This is a fundamental vulnerability! ➢ Major IM services do not obfuscate traffic patterns because it’s expensive 9

Our Attack A surveillance organization Adversary No need to cooperate with IM server Goal Identify participants of a target IM communication Timing Traffic Meta-data Identity of IM users Analysis Size 10

Attack Scenario Surveillance Area Adversary observes IM Server target user traffic A d v e r s a c r Target o y m o m b s u User e n r i v c e a s t i t g o a r n r o g t u e r n a t d f f t i c r u a t s h Target channel: “Let’s protest” 11

Adversary Ground Truth Adversary joins the Adversary observes target channel as a IM Server target user traffic member. A d v Target e Adversary joins the r s a c r o y User m o m b target channel as an s u e n r i v c e a s t i t g o a admin. r n r o g t u e r n a t d f f i t c r u a t s h Adversary wiretaps an identified Target channel: “Let’s protest” member/admin. 12

Target User Surveillance Area Target user is the Adversary observes admin of the target IM Server target user traffic communication A d v Target e r s a Target user is the c r o y User m o m b s u e n r member of the target i v c e a s t i t g o a r n r o g t communication u e r n a t d f f i t c r u a t s h Target channel: “Let’s protest” 13

Outline ❖ Modeling IM traffic: We established a statistical model for regular IM communications ❖ Design attack algorithms: We use hypothesis testing to design attack algorithms ❖ Experiments: We perform experiments on Telegram, WhatsApp, and Signal ❖ Countermeasures: We design and implement an open-source countermeasure system called IMProxy 14

Outline ❖ Modeling IM traffic: We established a statistical model for regular IM communications. ❖ Design attack algorithms: We use hypothesis testing to design attack algorithms ❖ Experiments: We perform experiments on Telegram, WhatsApp, and Signal ❖ Countermeasures: We design and implement an open-source countermeasure system called IMProxy 15

Modeling IM Traffic ❖ Deriving theoretical bounds on our traffic analysis algorithms. ❖ Generating synthetic IM communication. ❖ Dataset: Traffic patterns of 1000 Telegram channels, each for 24 hours. 16

Modeling IM Traffic IM Features Inter-Message Communication Message Types Message Sizes Delays (IMD) Latency 17

Outline ❖ Modeling IM traffic: We established a statistical model for regular IM communications ❖ Design attack algorithms: We use hypothesis testing to design attack algorithms ❖ Experiments: We perform experiments on Telegram, WhatsApp, and Signal ❖ Countermeasures: We design and implement an open-source countermeasure system called IMProxy 18

Attack Algorithms Event-Based Shape-based Algorithm Algorithm 19

Attack Algorithms: Event-Based If two events are close enough: 1- Event Extraction 2- Correlation Event MATCH! Function 3- Comparing to a Threshold Target User 20

Hypothesis Testing 21

Attack Algorithms: Shape-Based 1- Event Extraction 2- Traffic Normalization Cosine Traffic Bars Event 3- Correlation Similarity Function 4- Comparing to a Threshold Target User 22

Outline ❖ Modeling IM traffic: We established a statistical model for regular IM communications ❖ Design attack algorithms: We use hypothesis testing to design attack algorithms ❖ Experiments: We perform experiments on Telegram, WhatsApp, and Signal ❖ Countermeasures: We design and implement an open-source countermeasure system called IMProxy 23

Experimental Setup ❖ We perform experiments extensively on Telegram, WhatsApp, and Signal ❖ We use patterns of 500 channels. ❖ Scenarios ➢ Identifying Admin of a Telegram channel ➢ Wiretapping an identified user (one-to-one) 24

Attacks’ Performance Even with 15 minutes of traffic both algorithms have 94% Event-based algorithm Shape-based algorithm confidence while FP rate is 0.001 25

Why Not Deep Learning? We compared our work with DeepCorr We perform better than DeepCorr for smaller false positive rates!!? 1- IM flows are sparse. 2- IM flows are less noisy. 26

Outline ❖ Modeling IM traffic: We established a statistical model for regular IM communications ❖ Design attack algorithms: We use hypothesis testing to design attack algorithms ❖ Experiments: We perform experiments on Telegram, WhatsApp, and Signal ❖ Countermeasures: We design and implement an open-source countermeasure system called IMProxy 27

How to defend? 1- Using circumvention Event-based detector systems: Tor, VPN They are not effective without any background traffic. 28

IMProxy ❖ Algorithms: ❖ A proxy-based ➢ Adding delay obfuscation system ➢ Adding dummy ❖ No IM cooperation packets required ❖ Can be applied to any ❖ Main components: IM service just by ➢ Local proxy proxy the IM traffic ➢ Remote proxy through it 29

How It Works? Adversary Watching Adversary Watching IM Server Remote Remote proxy proxy Padding packets Removing padded packets Removing Padding packets padded packets Adding delay Local Local proxy proxy sender receiver Not observable by (admin) (member) adversary 30

Evaluating IMProxy ❖ Latency: A Laplacian distribution with parameter ❖ SOCKS5 proxy ❖ Adding dummy packets based on a Uniform Distribution ❖ Event-based attack With 10% bandwidth overhead, we have 30% decrease in confidence 31

Conclusions ❖ We show that despite the use of encryption, popular IM applications leak sensitive information about their client’s activities. ❖ The reason is that IMs do not use any obfuscation algorithms because it is expensive ❖ We hope that our results warn IM providers to take proper measures Thanks to 32

How It Works? S u r v e i l l a n c Remove Padded e A Padding packets packets r e a SIM Server Local proxy Remote proxy Padding packets and adding delays Remove Padded packets Remote proxy Local proxy 33

A Fundamental Vulnerability We show that despite the use of encryption, popular IM applications leak sensitive information about their client’s activities. Why? How? Merely watching Obfuscation of traffic encrypted IM traffic. is expensive for IM (Traffic Analysis) operators. 34

How to defend? 2- Using IMProxy Obfuscate timings by adding delays IMProxy: A proxy-based obfuscation system Obfuscate sizes by adding dummy traffic How it works? 35

Evaluating IMProxy ❖ Evaluating against IMProxy aware adversary ❖ Adversary trains a classifier on traffic flows 36

Attack Algorithms: Shape-Based Cosine Similarity Target User 37

Evaluating IMProxy ● Latency: A laplacian distribution with parameter ● Adding dummy packets based a Uniform Distribution ● SOCKS5 proxy Oblivious adversary IMProxy-aware adversary 38

Generalizing to other IMs Messages in IMs have the same shape of traffic They appear as bursts of packets Viber Signal WhatsApp Telegram 39

Telegram ❖ 200 million monthly active users. ❖ Most users are in countries with strict media regulations. Iran Russia ❖ It has the concept of channels. ❖ Telegram consumes 60! percent of Iran’s Internet bandwidth! 40