Unveiling the Digital Fingerprints: Analysis of Internet attacks based on website fingerprints

TL;DR

This paper demonstrates that advanced machine learning algorithms can effectively deanonymize Tor traffic by analyzing website fingerprints, revealing potential privacy vulnerabilities despite anonymization features.

Contribution

The study introduces a machine learning-based framework for website fingerprinting attacks on Tor, providing empirical evidence of their effectiveness with high accuracy.

Findings

Gradient Boosting Machine achieves 83.63% accuracy in binary classification.

Random Forest attains 62.97% accuracy in multi-class classification.

Machine learning can compromise Tor's anonymity protections.

Abstract

Parallel to our physical activities our virtual presence also leaves behind our unique digital fingerprints, while navigating on the Internet. These digital fingerprints have the potential to unveil users' activities encompassing browsing history, utilized applications, and even devices employed during these engagements. Many Internet users tend to use web browsers that provide the highest privacy protection and anonymization such as Tor. The success of such privacy protection depends on the Tor feature to anonymize end-user IP addresses and other metadata that constructs the website fingerprint. In this paper, we show that using the newest machine learning algorithms an attacker can deanonymize Tor traffic by applying such techniques. In our experimental framework, we establish a baseline and comparative reference point using a publicly available dataset from Universidad Del Cauca, Colombia. We capture network packets across 11 days, while users navigate specific web pages, recording data in .pcapng format through the Wireshark network capture tool. Excluding extraneous packets, we employ various machine learning algorithms in our analysis. The results show that the Gradient Boosting Machine algorithm delivers the best outcomes in binary classification, achieving an accuracy of 0.8363. In the realm of multi-class classification, the Random Forest algorithm attains an accuracy of 0.6297.