WF3A: A N-shot website fingerprinting with effective fusion feature attention

Abstract

Tor anonymity communication system provides privacy to users while also allowing sensitive users to access illegal websites. Website fingerprinting (WF) attack is a technique employed to identify websites visited by users, serving as a crucial tool for effectively recognizing user access to illegal online content. However, the majority of existing WF methods heavily rely on extensive training data, which faces challenges when dealing with more common few-shot scenarios which involve limited traffic traces. In such circumstance, these methods often fall into low accuracy in face of the absence of sufficient samples. Therefore, a N-shot WF recognition method with effective fusion feature attention (WF3A) is proposed in this paper. Different from the conventional single-feature recognition method, we innovatively combine direction and length, design and use the fusion feature that covers more website identification information; In order to interpret the traffic patterns that can identify the website categories from the fusion feature, we introduce the Enhanced Channel Attention (ECA) and design a stronger feature extractor when constructing WF model, which enhances the learning ability of the model by channel attention mechanism. The experimental results indicate that fusion feature outperforms original feature in few-shot scenarios. Furthermore, the proposed WF3A demonstrates superior performance compared to existing WF method. In the closed-world, the classification accuracy of WF3A is enhanced by 3.86% to 56.3% compared to existing methods. In the open-world, when the recall is controlled at 91.43%, the precision still remains stable at 81.16%.