Signal Chameleon: Robust Defense Against Tor Website Fingerprinting via Many-to-Many Traffic Morphing

Summary

Website fingerprinting attacks infer a user's browsing activity from encrypted Tor traffic by exploiting side-channel features. The authors found that most existing defenses still create learnable web trace mapping features, and they further showed that robustness against adversarial training does not guarantee robustness against defense-aware autoencoder based attacks. To address this they propose Chameleon, which selects morphing candidates with high intra-class diversity and low inter-class disparity, randomly maps each webpage trace to multiple candidates, and lets different webpages share morphing targets to raise an attacker's uncertainty. For practical Tor deployment Chameleon uses a radix-trie based synchronization mechanism so pluggable transport endpoints can identify consistent morphing traces from packet-direction prefixes alone. Evaluated against six defenses and five attacks on three public datasets, Chameleon cut adversarial-training based attack accuracy by up to 36 point 74 percent compared with prior defenses, while also reducing bandwidth and time overhead.