A First Look at Generating Website Fingerprinting Attacks via Neural Architecture Search

An adversary can use website fingerprinting (WF) attacks to breach the privacy of users who access the web through encrypted tunnels like Tor. These attacks have increasingly relied on the use of deep neural networks (DNNs) to build powerful classifiers that can match the traffic of a target user to the specific traffic pattern of a website. In this paper, we study whether the use of neural architecture search (NAS) techniques can provide adversaries with a systematic way to find improved DNNs to launch WF attacks. Concretely, we study the performance of the prominent AutoKeras NAS tool on the WF scenario, under a limited exploration budget, and analyze the effectiveness and efficiency of the resulting DNNs. Our evaluation reveals that AutoKeras's DNNs achieve a comparable accuracy to that of the state-of-the-art Tik-Tok attack on undefended Tor traffic, and obtain 5-8% accuracy improvements against the FRONT random padding defense, thus highlighting the potential of NAS techniques to enhance the effectiveness of WF.