EDF: Enhanced Deep Fingerprinting attacks on Websites

Website Fingerprinting (WF) attacks enable network surveillants to monitor traffic data and learn features from encrypted packet sequences, thus identifying which website user is visiting, which seriously threatens user privacy on Tor. Recently, zero-delay lightweight defense methods (i.e., FRONT and GLUE) were designed for Tor to efficiently defend many known WF attacks. But, can they defend more unknown intelligent WF attacks? Due to the efficiency of deep learning model, we explore whether there is a more intelligent deep learning model that could defeat FRONT and GLUE. In this work, we propose two deep learning based WF attacks, EDF-S and EDF-P. EDP-S incorporates a global average pooling layer and series model into Deep Fingerprinting (DF), while EDF-P introduces a global average pooling layer and parallel model into DF. We conduct extensive WF attack experiments on several datasets to prove our improvement on effectiveness. EDF-P can achieve over 97% TPR and 96% F1score on a non-defended dataset. We also find that EDF-S and EDF-P perform better than previous attacks against GLUE. Compared with DF, EDF-S can achieve almost 10% enhancement for the attack precision and F1score while EDF-P can achieve more than 10% enhancement for the attack precision and F1score. Our enhanced WF attacks bring new challenges to the above mentioned WF defenses.