AttCorr: A Novel Deep Learning Model for Flow Correlation Attacks on Tor

Abstract

The proliferation of illegal information and criminal behavior on anonymous networks arouses the demand for deanonymization attacks on anonymous communication systems. Flow correlation is a common technique for deanonymization attacks, but most related works currently are based on statistical learning models, thus suffering from burdensome artificial feature engineering. In this paper, we design a novel deep learning model AttCorr for flow correlation attacks on Tor. AttCorr takes the raw traffic features of flow pair as input, including packet sizes, flow directions and inter-packet delays, and uses the multi-head attention mechanism to capture the flow information involved in features and the complex nature of noise in Tor. Experiment results show that AttCorr achieves the same level of accuracy as the stateof-the art method of DeepCorr with lower complexity, simpler feature processing and better interpretability.