Revisiting Website Fingerprinting Attacks in Real-World Scenarios: A Case Study of Shadowsocks

Abstract

Website fingerprinting has been recognized as a traffic analysis attack against encrypted traffic induced by anonymity networks (e.g., Tor) and encrypted proxies. Recent studies have demonstrated that, leveraging machine learning techniques and numerous side-channel traffic features, website fingerprinting is effective in inferring which website a user is visiting via anonymity networks and encrypted proxies. In this paper, we concentrate on Shadowsocks, an encrypted proxy widely used to evade Internet censorship, and we are interested in to what extent state-of-the-art website fingerprinting techniques can break the privacy of Shadowsocks users in real-world scenarios. By design, Shadowsocks does not deploy any timing-based or packet size-based defenses like Tor. Therefore, we expect that website fingerprinting could achieve better attack performance against Shadowsocks compared to Tor. However, after deploying Shadowsocks with more than 20 active users and collecting 30 GB traces during one month, our observation is counter-intuitive. That is, the attack performance against Shadowsocks is even worse than that against Tor (based on public Tor traces). Motivated by such an observation, we investigate a series of practical factors affecting website fingerprinting, such as data labeling, feature selection, and number of instances per class. Our study reveals that state-of-the-art website fingerprinting techniques may not be effective in real-world scenarios, even in the face of Shadowsocks which does not deploy typical defenses.