Adversarial Vulnerability Under Temporal Concept Drift: A Longitudinal Study of Android Malware Detection
Researchers have conducted a longitudinal study on the adversarial robustness of Android malware detection systems over a decade. Their findings indicate that temporal concept drift, where data distributions change over time, significantly reduces the adversarial robustness of these systems. While retraining models with cumulative historical data can mitigate some of this loss, it does not entirely eliminate the problem, highlighting the need for drift-aware assessment frameworks. AI
IMPACT Highlights the need for drift-aware robustness assessment in long-lived adversarial AI systems, impacting developers of security AI.