New Method Detects Adversarial Attacks Using Uncertainty Patterns

A new research paper introduces USAD, or Uncertainty-aware Statistical Adversarial Detection, a novel approach to identifying adversarial examples (AEs) by focusing on their unique uncertainty characteristics. Traditional statistical adversarial detection (SAD) methods often rely on measuring general distributional discrepancies, which can overlook crucial patterns specific to AEs. USAD addresses this limitation by explicitly modeling how AEs exhibit abnormal feature spread and instability under minor perturbations. The core of USAD involves two new statistical measures: Variance Discrepancy (VD) and Perturbation-based Covariance Discrepancy (PCD). VD quantifies differences in feature spread, capturing global uncertainty, while PCD compares feature covariance under Gaussian perturbations to identify local uncertainty. By combining these two metrics, USAD significantly outperforms existing SAD methods in detecting a wide range of adversarial attacks, underscoring the importance of understanding the distinct behaviors of AEs for robust detection.