A Practical Website Fingerprinting Defense Approach with Universal Adversarial Perturbations

Abstract

Website Fingerprinting (WF) can be used by network eavesdroppers to infer information about the content of encrypted and anonymous connections. Several defenses leverage adversarial examples to mitigate the threat of WF attacks, but they require to get entire traffic traces after network sessions have concluded to produce adversarial examples, thus offer users little protection in practical settings. In this paper, a novel practical WF defense method called WF-UAP is proposed by crafting Universal Adversarial Perturbations (UAP) to fight back against WF attacks, in which adversarial examples are produced to fool the classifiers used in WF attacks when UAPs are added to any network traffic trace in the target domain. We further present a Generative Adversarial Network (GAN) based UAP generation approach to enhance the performance of UAPs. It can efficiently generate UAPs once the generator is trained and make the adversarial traces and original traffic traces are more difficult to distinguish. Our experimental results over a public dataset demonstrate that WF-UAP reduces the accuracy of the state-of-the-art WF attacks from 98% to 15% with at most 20% increased bandwidth overhead, which outperforms the previous defenses in terms of the defense performance and overhead.