A Comparative Analysis of Feature Representation Paradigms for Website Fingerprinting on the Java Anon Proxy Network

Website fingerprinting (WF) attacks pose a serious threat to online privacy, even when users employ anonymizing networks. The effectiveness of deep learning-based WF critically depends on how raw network telemetry is transformed into input representations. In this paper, we systematically compare five feature encoding paradigms–Direction Only, Inter-arrival Times, Unsigned Packet Sizes, Signed Inter-arrival Times, and Signed Packet Sizes–using two deep learning architectures: a 1D Convolutional Neural Network (CNN) and a CNN-LSTM-Attention (CLA) model. We conduct our experiments on a newly collected Java Anon Proxy (JAP) dataset of 10,000 traces, measuring accuracy, stability (standard deviation and learning curves), training time, and peak RAM usage. Our results show that Direction Only achieves high accuracy (99.14% with CLA) and low variance, confirming its discriminative power. However, Signed Packet Sizes combining packet size with directional sign–deliver the best overall performance: CLA reaches 99.56% accuracy with low variance (0.24%), while the CNN model attains 98.22% with even lower variance (0.18%) and faster training time. A sensitivity analysis with four weighting schemes (accuracy-focused, efficiency-focused, stability-focused, balanced) shows that while Signed Sizes CLA is optimal for absolute peak accuracy, Signed Sizes-CNN provides the best trade-off for practical deployments where computational resources are limited.