ANALYZING DARKNET TRAFFIC: IMPACT OF MODIFIED TOR TRAFFIC PATTERNS ON ONION SERVICE CLASSIFICATION ACCURACY

Abstract

Darknet communication, especially through the Tor network, has become a crucial focus of cybersecurity research due to increasing misuse by malicious actors. However, identifying and classifying Onion Service traffic remains challenging because Tor encrypts packet contents and introduces obfuscation mechanisms. This study investigates how modified Tor traffic patterns, including padding variations, timing perturbations, and circuit-level manipulation, influence the accuracy of Onion Service traffic classification systems. A hybrid analytical model combining flow-level statistical features, deep learning– based traffic embeddings, and adversarial perturbation analysis is used to evaluate classification resilience. Experimental results show that even minor manipulations significantly reduce classifier confidence and increase misclassification rates, emphasizing the vulnerability of current Darknet monitoring systems [1], [3], [5], [9]. The findings provide critical insights for designing more robust traffic analysis systems capable of detecting evasion strategies within Tor-based communication.