DeTorrent: An Adversarial Padding-only Traffic Analysis Defense

Anonymity networks like Tor remain vulnerable to traffic-analysis attacks such as Website Fingerprinting (WF) and Flow Correlation (FC), and recent attacks like Tik-Tok and DeepCoFFEA have made these threats increasingly practical. DeTorrent uses competing neural networks (an adversarial generator-and-evaluator design) to produce padding-only defenses that insert dummy traffic into real flows without delaying user packets. In a closed-world WF setting it reduces an attacker's accuracy by 61.5% — 10.5 points better than the next-best padding-only defense — and against the state-of-the-art FC attacker it cuts the true-positive rate at a 1e-5 false-positive rate to roughly half that of the next-best defense. The authors also deploy it alongside live Tor traffic to demonstrate its practicality.