MW3F: Improved multi-tab website fingerprinting attacks with Transformer-based feature fusion

Abstract

Website Fingerprinting (WF) attacks compromise the anonymity of Tor by analyzing traffic patterns. Multi-tab WF attacks, which aim to identify multiple categories of websites from obfuscated traffic, have achieved significant progress. However, existing methods often fail to fully exploit the relationships between traffic features. On the one hand, splitting-based methods have complex processes that result in the loss of local traffic features. On the other hand, end-to-end methods process complete traffic but perform poorly when relying on a single feature. To address these challenges, this paper proposes an effective Multi-tab Website Fingerprinting attack with Transformer-based Feature Fusion named MW3F. Specifically, MW3F first extracts high-level traffic features, including direction and inter-packet time. Subsequently, These new representations are fused using the multi-head self-attention, which captures both local dependencies and global interactions. Finally, to identify website categories adaptively, MW3F incorporates learnable label embeddings to probe and pool class-related features. Each website category prediction is associated with a corresponding label embedding. We evalute MW3F against state-of-the-art multi-tab WF attacks in both multi-tab and defense scenarios. In the closed-world scenario, MW3F achieves a mean average precision (mAP) of over 90% across all tab settings, outperforming the strongest baseline, ARES, by 7% in the 5-tab setting. In the defense scenario, MW3F achieves approximately 90% mAP against WTF-PAD and Front defenses, demonstrating superior performance and exceptional robustness.