TL;DR
This paper introduces a novel traffic confirmation attack that uses robust, real-time binary hashes of network traffic flows to link senders and receivers in low-latency anonymity networks like Tor. The attack is highly accurate even with packet loss and noise, and it demonstrates that even advanced probabilistic padding schemes are vulnerable to this type of deanonymization.
What Problem Does This Paper Address?
Anonymity networks like Tor aim to obscure the origin and destination of internet traffic by routing it through multiple relays. While encryption hides the content, metadata like packet timing and size can still reveal patterns. Traffic confirmation attacks exploit these patterns to link traffic at the network's entry to traffic at its exit, thereby deanonymizing users. Previous attacks faced challenges due to network noise, packet loss, and the computational cost of matching large amounts of data. This paper addresses the question of whether accurate traffic confirmation attacks are still possible in the presence of these real-world network impairments and against sophisticated defenses like probabilistic padding.
Methods
The core of the attack relies on a modified version of the Coskun-Memon Algorithm to generate "robust real-time binary hashes" of network traffic flows. Instead of directly comparing raw packet timing data, the algorithm first divides traffic into time windows and counts packets within each. It then computes the difference in packet counts between consecutive windows and uses pseudo-random bases to project this information into a compact binary hash. This hashing approach is designed to be computationally efficient and resilient to network impairments like packet loss and timing variations. The authors then use these hashes to match network flows observed at different points in the network. They also introduce an adaptation of this method, "adapted-SCC," specifically to counter probabilistic padding schemes.
Datasets
The primary dataset used is the publicly available Wang et al. dataset [22], which contains 90 instances of 100 sensitive websites accessed via Tor. To simulate real-world network conditions, the authors modified this dataset to include:
- Packet drop probabilities: 1%, 5%, 10%, and 30%.
- Timestamp obfuscation: Adding random delays drawn from an experimentally derived Inter-Packet Delay Variation (IPDV) distribution.
They also generated their own IPDV distribution through real-world experiments on the Tor network to better model network impairments.
Key Insights and Ideas
- Hashing for Robustness: Using binary hashes of traffic flow characteristics, rather than raw timing data, significantly improves the attack's resilience to noise and packet loss. This makes the attack more practical in real-world scenarios.
- Low Cost, High Accuracy: The hashing approach is computationally efficient for both generating and matching hashes, making it feasible for large-scale attacks. The attack achieves high accuracy even when only a fraction of the traffic is observed.
- Padding is Not a Silver Bullet: Advanced probabilistic padding schemes, designed to thwart timing attacks, are still vulnerable. The adapted-SCC algorithm demonstrates that by analyzing the structure of padded flows relative to unpadded ones, deanonymization remains possible.
- Tor's Vulnerability: The paper specifically targets Tor, the most popular low-latency anonymity network, highlighting its susceptibility to these refined traffic confirmation attacks.
Key Results
- Unpadded Mixes: The attack achieves an 80% true positive match rate when identifying one flow out of 9000, with less than a 2% false positive rate, even with up to 10% packet drop.
- Website Fingerprinting: The attack can successfully link flows to the correct website with over 80% accuracy, degrading to 72% when monitoring 100 websites.
- Padded Mixes (Adaptive Padding): The adapted-SCC algorithm achieves over 90% match rate for padded flows when the IPDV average is 21ms (representative of real Tor conditions). The attack remains effective even with higher IPDV averages.
Limitations and Future Work
- Live Tor Implementation: The authors acknowledge that a live implementation on Tor would require careful handling to avoid capturing other users' traffic while still ensuring background traffic is present.
- Padding Schemes: While the attack works against Adaptive Padding (AP), the authors suggest evaluating other padding schemes, such as dependent link padding or intentional packet delays, as future work.
- Real-World Network Dynamics: The current IPDV simulation is based on independent packet delays. Future work could explore more realistic scenarios where packet delays are dependent on network load.
- Protocol Changes for Padding: The paper notes that for AP to be fully effective in Tor, protocol changes might be necessary to allow relays to generate dummy cells, which is currently not feasible due to Tor's integrity checks.
Summary updated