Neural Architecture Search for Website Fingerprinting

Abstract

To protect their online privacy, users often employ encrypted tunnels from networks like Tor, which create multi-hop pathways to conceal communications. Tor’s design, while effective, preserves packet timing and volume, making it vulnerable to website fingerprinting (WF) attacks. In these attacks, eavesdroppers use machine learning to match traffic patterns to specific websites. Recent WF research has shifted from traditional machine learning to deep neural networks (DNNs) for more precise attacks. This shift has led to a reliance on trial-and-error adaptations from other domains, potentially overlooking innovative architectural choices.

To address this challenge, this thesis introduces FAWNS, a NAS framework for website fingerprinting and traffic analysis, integrated with Microsoft’s NNI AutoML toolkit. FAWNS automates DNN design using a predefined search space. Consequently, our evaluation shows that FAWNS-generated DNNs achieve accuracy comparable to state-of-the-art attacks on undefended Tor traffic and significantly higher accuracy on defended traffic, with a 26% increase on FRONT-protected traces and a 12% increase on WTF-PAD protected traces. These models also generalize well to unseen data, highlighting NAS's potential to enhance WF effectiveness.