A WASSERSTEIN GAN BASED FRAMEWORK FOR ADVERSARIAL ATTACKS AGAINST INTRUSION DETECTION SYSTEMS

Abstract

IIntrusion detection system (IDS) detects malicious activities in network flows and is essential for modern communication networks. Machine learning (ML) and deep learning (DL) have been employed to construct IDSs. However, the reliability of ML/DL-based IDSs is questionable under adversarial attacks. We propose a framework based on Wasserstein generative adversarial networks (WGANs) to generate adversarial traffic to evade ML/DL-based IDSs. The proposed framework involves restricted modification operations and the output of the framework is carefully regulated, preserving the functionality of the malicious attack. We also present a variant of the proposed framework based on conditional WGANs. The variant framework simplifies the training procedure without losing attack capability. Eight ML/DL-based IDSs are constructed, and their robustness against adversarial attacks is tested using the frameworks. The results show that the framework and its variant can generate adversaries effectively, and the Convolutional Neural Network has the best robustness under adversarial attacks.