A Multi-Granularity Features Representation and Dimensionality Reduction Network for Website Fingerprinting

Website fingerprinting (WF) refers to the identification of target websites accessed by users in anonymous communication scenarios, playing a critical role in cybercrime investigation and forensics. In recent years, deep learning-based WF has become a research focus in this field. Although these methods can achieve high accuracy, their performance tends to decline over time. This degradation results from the evolution of network protocol versions and the ongoing development of obfuscation techniques, a phenomenon known as concept drift. To address the problem of concept drift, this paper presents a multi-granularity features representation and dimensionality reduction network for Website Fingerprinting, referred to as LRCT. The LRCT network effectively leverages the temporal learning advantages of Local Recurrent Networks (Local RNN) and the spatial learning strengths of Convolutional Neural Network (CNN) by designing the local feature extraction block (denoted as LRC Block), which extracts fine-grained local features from 2000-dimensional original sequences and reduces the dimensionality to 125. The network then uses a Transformer Encoder to capture more robust global features from the low-dimensional data. Experimental results comparing LRCT with state-of-the-art WF methods demonstrate that the proposed LRCT achieves 99.34% accuracy in a closed-world scenario, outperforming other benchmark models. Compared to the benchmark WF-Transformer method, the LRCT network reduces the parameter count by half and decreases the average training time per epoch to approximately 20% of the benchmark. Even under concept drift conditions, LRCT maintains an accuracy of 89.9%. In open-world scenarios, LRCT also has better precision and recall.