ProxyCorr: robust traffic correlation attacks via mixed spatio-temporal analysis in encrypted proxy networks

Abstract

End-to-end traffic correlation attacks, aiming to deanonymize users in anonymous communication systems, have achieved significant advancements through deep learning. However, existing methods suffer significant limitations in encrypted proxy networks, especially when obfuscation techniques are employed. To address these limitations, we propose ProxyCorr, a model that achieves accurate traffic correlation across diverse encrypted proxy protocols. Specifically, we model network flows as state sequences and derive temporal dependencies through state transition analysis. Additionally, to capture spatial similarity patterns, we model flows as 2D spatial trajectories, with Gaussian filtering applied to mitigate jitter. Further, we design a self-attention-based correlation learning module to extract inherent spatio-temporal features, which are then adaptively integrated to enhance flow correlation performance. Extensive experiments on multiple encrypted-proxy datasets demonstrate that ProxyCorr outperforms state-of-the-art methods in effectiveness while maintaining robustness against obfuscation techniques, including packet size padding and multiplexing.