Securing Medical AI: Adversarial Attack Detection and Mitigation in CAD Systems

Abstract

Frequency-constrained adversarial attacks introduce imperceptible, high-frequency distortions that undermine the reliability and safety of AI-assisted medical diagnostics, posing a serious challenge to trust and resilience in socially-interactive, human-AI healthcare systems. To address this, we introduce Sentinel, a pioneering framework that detects, classifies, and mitigates adversarial attacks in medical image analysis without requiring training or model specificity. Sentinel integrates a lightweight detection module that uses a confidence scorer to analyze predictions from multiple low-rank approximations of the input image, enabling precise classification of attacks as either gradient-based or frequency-constrained. Uniquely, Sentinel includes a recovery module specifically designed for frequency-constrained attacks, employing adaptive Robust Principal Component Analysis-based denoising to suppress perturbations concentrated in smaller singular values while preserving essential image structure. This allows for effective image reconstruction, an area overlooked by existing methods. Evaluated across four diverse medical imaging datasets, Sentinel demonstrates consistently high detection accuracy, classification performance, and adversarial recovery without the need for model retraining or GPU support, making it a highly practical and scalable solution for integration into clinical AI systems. Additional explainability analyses confirm the reliability and interpretability of Sentinel's predictions, establishing it as a significant advancement in adversarial defense.