Adversarially Robust Few-Shot Anomaly Detection with Vision Foundation Models
The study introduces a robust few-shot anomaly detection method using frozen Vision Foundation Models. It solves the problem of attacking non-parametric scorers by creating a differentiable surrogate called DistanceProbe. This enables strong gradient-based attacks, and it uses a two-level defense (PatchShift and FeaturePurifier) to significantly boost robustness against adversarial attacks while preserving clean performance.