Keyed randomization with adversarial failure curves and moving target defense

Establishing the robustness of classifiers against adversarial attacks is crucial in many applications of Machine Learning to Cybersecurity. This paper focuses on evasion attacks, where inputs are selected or modified to evade detection by a learned model under gray-box scenarios, with only partial adversarial knowledge of the classifier. We generalize the adversarial failure rate metric into a continuous curve, by trading it off against the false positive rate for threshold classifiers, analogous to the receiver operating characteristic (ROC) curve. Subsequently, we propose two novel keyed randomization methods, and a moving target defense strategy. We evaluate the proposed methods using two publicly available intrusion detection datasets (BETH-2021 and Kyoto-2015), demonstrating consistently superior results relative to other randomization techniques.