U-ASG: A Universal Method to Perform Adversarial Attack on Autoencoder based Network Anomaly Detection Systems

摘要

Semi-supervised machine learning models, especially deep neural networks, have been widely used in network anomaly detection for their capability of capturing patterns in normal data. However, the models face security challenges when an attacker has obtained their full details. In this paper, we propose a universal adversarial sample generator (U-ASG), to perform white-box adversarial attacks on autoencoder-based semi-supervised network anomaly detection (SSNAD) systems. The purpose of adversarial attacks is to generate small adversarial perturbations and add them to targeted anomalous samples to fly under the radar. We model the generation process of adversarial perturbations as an optimization problem, in which we minimize the reconstruction errors of the adversarial samples through the trained autoencoder and approximate it to solve. Furthermore, to improve the attack performance against the variational autoencoder (VAE), which is robust to tiny perturbations through uncertainty modeling, we design a mechanism to weaken its robustness by introducing a variance regularizer to the optimization. Simulation results show that the adversarial attacks generated by our U-ASG can effectively degrade the performance of the autoencoder-based SSNAD systems.