Robust Website Fingerprinting Via Multi-Layer Feature Representation and Adaptive Channel Attention

The Tor anonymity network, while protecting user privacy and circumventing internet censorship, remains vulnerable to website fingerprinting (WF) attacks that correlate users with visited websites to breach privacy. Although multiple defense mechanisms exist against WF attacks, achieving more accurate and robust WF attacks remains critical. This paper proposes a novel WF attack named Multilayer Attention Fingerprinting (MAF), which employs a Multilayer Analysis Matrix (MAM) containing richer latent features to comprehensively characterize traffic distinctiveness. The MAF framework trains a channel attention model with dynamic learning rates to integrate global feature correlations. Extensive experiments on public datasets-including open-world, closed-world scenarios, and datasets protected by various WF defenses-demonstrate MAF's superior performance. It achieved 99% accuracy against Walkie-Talkie defense-enhanced datasets, and 98% accuracy on both closed-world and open-world datasets. These results validate MAF's high accuracy and robustness against existing WF defenses.