Despite the use of encryption, the low-latency nature of the Tor anonymity network leaves users’ traffic exposed to traffic analysis attacks aimed at uncovering the websites visited by users. These attacks, known as website fingerprinting (WF), have increasingly relied on deep neural networks (DNNs) due to their effectiveness in identifying complex traffic patterns. Thus far, however, the development of suitable DNN architectures for WF attacks has resorted to trial-and-error, potentially overlooking more effective and/or efficient attack models. In this paper, we introduce Cricket, a framework for the automatic exploration of DNNs tailored to complex WF tasks. Cricket leverages and adapts DARTS, a prominent neural architecture search (NAS) approach, to discover new DNN architectures that perform best for a given WF task. In evaluating Cricket, we find that: (1) in a closed-world setting, Cricket consistently outperforms the SOTA attacks by 2–7 points and up to 13 points against WF defenses; (2) in an open-world setting, Cricket offers comparable performance to SOTA attacks against SOTA defenses; and (3) the novel DNNs produced by Cricket use 1/3 to 1/2 as many parameters as the SOTA models. Our results show how Cricket adapts to the WF context to produce novel DNNs that are both more effective and more compact than existing models.
@inproceedings{cricket-wpes2026,
title = {Neural Architecture Search for Adaptive Website Fingerprinting},
author = {Singh, Gurjot and Jansen, Rob and Barradas, Diogo},
booktitle = {Workshop on Privacy in the Electronic Society},
year = {2026},
doi = {10.1145/3847192.3847366},
}