DTLF: Deep Transfer Learning for Website Fingerprinting

Abstract

In anonymous networks, users can be tracked by implementing website fingerprinting (WF) attacks, which can prevent users from browsing illegal websites or performing illegal actions. With deep learning technologies, WF attacks achieve higher than 98% accuracy against powerful anonymity Tor network. However, these attacks rely on enormous encrypted traffic data, which are time-consuming to collect. Moreover, the large-scale encrypted traffic data collected also needs to be updated frequently to adapt for changes in website content. This paper proposes a new method, Deep Transfer Learning For Website Fingerprinting (DTLF), which combines the deep transfer learning framework with model DResNet. This method obtains useful and rich feature representations by training on a large number of existing labeled datasets to apply to few encrypted traffic. The results over multiple datasets show that the DTLF can achieve 96.5% accuracy with 5 labeled samples. In the 56-day experimental setting against concept drift, the accuracy is only 2.6% lower. In addition, the DTLF can consume less time overhead to perform WF attack. When training on the target dataset, it often only takes tens of seconds which is less than 0.1 times of other methods.