On the Application of Deep Learning Techniques to Website Fingerprinting Attacks and Defenses

Several studies have shown that traffic metadata can be exploited by a network-level adversary to identify the websites that users are visiting over Tor. The success of such attacks, known as Website Fingerprinting attacks, heavily depends on the particular set of traffic features that are used to distinguish websites. Typically, these features are manually engineered and static which makes them fragile to changes in the Tor protocol and the deployment of defenses. In this work we evaluate a traffic analysis attack based on deep learning techniques that allows us to extract features automatically. We show that our attack’s performance is comparable to that of traditional attacks, while eliminating the need for feature design and selection. We argue that this may be a game-changer in the arms-race between Website Fingerprinting attacks and defenses.