Adversarial Evaluation of AI-Based Security Alert Screening Systems
Samuel Ndichu, Tao Ban, Takeshi Takahashi, Akira Yamada, Seiichi Ozawa, Daisuke Inoue · 2024
Artificial intelligence (AI)-based security alert screening systems are vulnerable to adversarial attacks that cleverly manipulate network traffic data to evade detection. This study evaluated the robustness of AI-based security alert screening systems against practical black-box attacks in a controlled environment. Using the Tcpreplay toolbox, we manipulated a pre-existing security alert log dataset to emulate realistic adversarial activities, including normal traffic flooding, distributed denial-of-service attacks, injection of fake critical alert logs, and a combination of these scenarios. We propose smart sample selection techniques to identify ambiguous samples for adversarial manipulation applying a threshold-based approach to the probabilities predicted by the classifier. Our findings from the evaluation of five AI-based security alert screening system implementations revealed significant vulnerabilities in the ability of the system to discern manipulated traffic, underscoring the need for improved detection algorithms that can withstand sophisticated adversarial strategies. This study contributes to the cybersecurity field by highlighting the critical areas in which current AI-based security alert screening systems can be enhanced to counteract emerging adversarial strategies.