A Label-Level Triplet Network for Multi-Label Website Fingerprinting Attacks

Website fingerprinting attacks (WFAs) constitute a significant threat to user privacy by exploiting side-channel information from encrypted communications. However, existing methods often suffer from limited recognition accuracy and constrained feature representation, particularly in the context of multi-label complexities and diverse attack patterns. To address these challenges, we propose a Label-Level Triplet Network (LLTN) designed for multi-label scenarios. This method incorporates label information into the triplet construction process, establishing cross-label hierarchical structural constraints that enable fine-grained modeling of semantic similarities and discriminative features across multidimensional attack patterns. By explicitly optimizing label-driven inter-sample distances in the embedding space, LLTN significantly enhances discriminability and generalization performance under complex multi-label conditions. Moreover, the framework effectively addresses label imbalance, ensuring robust model convergence and improved classification stability. Extensive experiments on multiple real-world WFA datasets demonstrate that the proposed method consistently outperforms state-of-the-art approaches, particularly in high label-overlap and low signal-to-noise ratio scenarios. Overall, LLTN offers a promising solution for accurate modeling and detection of multi-label WFAs, while also extending the application scope of triplet networks in network traffic analysis and privacy-preserving technologies.