TrafficSiam: More Realistic Few-shot Website Fingerprinting Attack with Contrastive Learning

Website fingerprinting (WF) attacks pose a serious threat to users’ online privacy, even when using privacy-enhancing tools like Tor. Previous attacks mostly rely on supervised learning and only a few studies have explored the few-shot setting which is more realistic. In this paper, we propose TrafficSiam, a novel WF attack based on few-shot learning with self-supervised learning, which enables our model to be pretrained with unlabeled tor traffic and transferred to new tasks using a small number of labeled samples which is more realistic and yield stronger generalization ability. We conduct a series of experiments, using only a small amount of labeled samples, and find that our model achieves 92.32% accuracy in the closed-world setting, compared to the highest accuracy 88.73%, using previous methods. Furthermore, our model also outperforms previous attacks in the open-world setting.