Empirical Analysis of Adversarial Robustness and Explainability Drift in Cybersecurity Classifiers

AI security tools can be fooled by tiny, intentional changes: The study found that phishing detection and network intrusion systems—which organizations rely on to catch threats—can be degraded by small adversarial perturbations, meaning attackers could potentially craft inputs that bypass these AI defenses.

When AI models are attacked, they become harder to understand: The research reveals "explainability drift," meaning that as adversarial attacks degrade model performance, the explanations for why the AI made its decisions also become unreliable. This creates a double problem: the system fails and you can't trust why it failed.

Adversarial training offers meaningful but incomplete protection: Adversarial training improved the Robustness Index by up to 9 percent while maintaining accuracy on normal data, suggesting this is a practical defense strategy—but the modest improvement indicates it's not a complete solution to the vulnerability problem.

Certain features are more vulnerable than others: By analyzing which input features are most susceptible to manipulation (using gradient-based and SHAP-based methods), the study identifies specific weaknesses in how these systems process cybersecurity data, enabling more targeted hardening efforts.

Quantitative robustness metrics are essential for trustworthy AI deployment: The paper emphasizes that organizations deploying AI for cybersecurity need standardized, measurable ways to evaluate robustness—not just accuracy—to ensure these systems are genuinely reliable before protecting critical infrastructure.