Discovering Imperfectly Observable Adversarial Actions using Anomaly Detection

Defenders in security problems often use anomaly detection (AD) to examine effects of (adversarial) actions and detect malicious behavior. Attackers seek to accomplish their goal (e.g., exfiltrate data) while avoiding the detection. Game theory can be used to reason about this interaction. While AD has been used in game-theoretic frameworks before, we extend the existing works to more realistic settings by (1) allowing players to have continuous action spaces and (2) assuming that the defender cannot perfectly observe the action of the attacker. We solve our model by (1) extending existing algorithms that discretize the action spaces and use linear programming and (2) by training a neural network using an algorithm based on exploitability descent, termed EDA. While both algorithms are applicable for low feature-space dimensions, EDA produces less exploitable strategies and scales to higher dimensions. In a data exfiltration scenario, EDA outperforms a range of classifiers when facing a targeted exploitative attacker.