CTWF: Website Fingerprinting Attack based on Compact Convolutional Transformer

Website fingerprinting (WF) aims to identify the specific websites by analyzing traffic data patterns. Deep convolutional neural networks (CNNs) currently lead in WF attack methodologies for the identification of anonymous network traffic. However, CNN architectures underperform in capturing long-term temporal dependencies. This paper investigates methods to enhance the effectiveness of website fingerprinting attacks by improving the model while striving to conserve computational resources. We introduce a Compact Convolutional Transformer Website Fingerprinting (CTWF) attack model. The CTWF model is specifically designed for Tor anonymous traffic, fully leveraging the Transformer to capture temporal dependencies and utilizing the sequence of Tor cell bursts to compress input. It features a streamlined architecture with convolutional token extraction layers and sequence pooling layers for efficient attack modeling. We evaluated the effectiveness of the CTWF attack in both closed-world and open-world scenarios. Our results show that CTWF significantly outperforms the most advanced attack models, with and without time drift and defenses.