FSTC: A Few-Shot Website Fingerprinting Attack via Trace Augmentation and Contrastive Learning

Anonymity networks protect private communication by hiding users' online destinations. However, this anonymity also makes malicious online activities harder to analyze. Website fingerprinting is used to determine which websites a user is visiting over an encrypted connection. Although website fingerprinting attacks have achieved promising results by leveraging deep learning, they typically depend on large labeled datasets and often suffer performance degradation under changing network conditions. We propose a few-shot website fingerprinting attack (named FSTC) to alleviate these challenges. We augment traces to simulate network variations and use contrastive learning to learn discriminative representations from only a few labeled samples. Furthermore, we design a precision optimization mechanism to improve the effectiveness of the attack in open-world settings. Extensive open-world and closed-world experiments show that our attack performs effectively. In the 1-shot closed-world setting, our method achieves an accuracy of 76.4%, outperforming the compared methods (e.g., Sawllow) by approximately 10%. In the open-world setting, FSTC attains a recall of 89.1% and a precision of 89.2%, which demonstrates the strong capability of our attack to distinguish monitored traffic from complex background traffic. These results show that FSTC provides an effective solution for few-shot website fingerprinting attacks under dynamic network environments.