Adversarial Pre-Padding: Generating Evasive Network Traffic Against Transformer-Based Classifiers

TL;DR

This paper introduces AdvTraffic, a novel adversarial traffic generation method using pre-padding and reinforcement learning to evade transformer-based network traffic classifiers, significantly reducing their accuracy.

Contribution

It presents the first adversarial perturbation technique specifically designed to counter transformer-based traffic classifiers, with practical deployment considerations.

Findings

Reduces classifier accuracy from 99% to 25.68%

Effective against multiple real-world datasets

Outperforms existing obfuscation methods

Abstract

To date, traffic obfuscation techniques have been widely adopted to protect network data privacy and security by obscuring the true patterns of traffic. Nevertheless, as the pre-trained models emerge, especially transformer-based classifiers, existing traffic obfuscation methods become increasingly vulnerable, as witnessed by current studies reporting the traffic classification accuracy up to 99\% or higher. To counter such high-performance transformer-based classification models, we in this paper propose a novel and effective \underline{adv}ersarial \underline{traffic}-generating approach (AdvTraffic\footnote{The code and data are available at: https://anonymous.4open.science/r/TrafficD-C461}). Our approach has two key innovations: (i) a pre-padding strategy is proposed to modify packets, which effectively overcomes the limitations of existing research against transformer-based models for network traffic classification; and (ii) a reinforcement learning model is employed to optimize network traffic perturbations, aiming to maximize adversarial effectiveness against transformer-based classification models. To the best of our knowledge, this is the first attempt to apply adversarial perturbation techniques to defend against transformer-based traffic classifiers. Furthermore, our method can be easily deployed into practical network environments. Finally, multi-faceted experiments are conducted across several real-world datasets, and the experimental results demonstrate that our proposed method can effectively undermine transformer-based classifiers, significantly reducing classification accuracy from 99\% to as low as 25.68\%.