Mockingbird: Defending Against Deep-Learning-Based Website Fingerprinting Attacks With Adversarial Traces

Website Fingerprinting (WF) is a type of traffic analysis attack that enables local passive eavesdropper to infer the victim's activity, even when protected by VPN or an anonymity system like Tor. Leveraging deep-learning classifier, WF attacker can gain over 98% accuracy on Tor traffic. In this paper, we explore novel defense, Mockingbird, based idea adversarial examples have been shown undermine machine-learning classifiers in other domains. Since gets design and train his classifier first demonstrate at straightforward technique for generating adversarial-example traces fails protect against using training robust classification. We then propose resists moving randomly space viable not following more predictable gradients. The drops state-of-the-art hardened with from 42-58% while incurring only 58% bandwidth overhead. generally lower than defenses, much considering Top-2 accuracy, overheads.