Adversarial attacks based on time-series features for traffic detection

Abstract

To enhance the robustness of intrusion detection classifiers, we propose a Time Series-based Adversarial Attack Framework (TSAF) targeting the temporal characteristics of network traffic. Initially, adversarial samples are generated using the gradient calculations of CNNs, with updates iterated based on model loss. Different attack schemes are then applied to various traffic types and saved as generic adversarial perturbations. These time series-based perturbations are subsequently injected into the traffic stream. To precisely implement the adversarial perturbations, a masking mechanism is utilized. Our adversarial sample model was evaluated, and the results indicate that our samples can reduce the accuracy and recall rates for detecting four types of malicious network traffic, including botnets, brute force, port scanning, and web attacks, as well as degrade the detection performance of DDoS traffic. The CNN model’s accuracy dropped by up to 72.76%, and the SDAE model’s accuracy by up to 78.77% with minimal perturbations. Our adversarial sample attack offers a new perspective in the field of cybersecurity and lays the groundwork for designing AI models that can resist adversarial attacks more effectively.