Deep learning systems detect hidden attacks on medical images
Robust Detection of Directional Adversarial Attacks in Deep Neural Networks for Radiological Imaging
Computer Vision and Pattern RecognitionCryptography and SecurityMachine Learning
Summary
Deep learning helps doctors analyze medical scans, but small subtle changes that fool AI systems can cause errors. The authors propose a way to spot these sneaky tricks by adding random noise and checking how the AI’s predictions change. Their method can tell real medical images from attacked ones with high accuracy, keeping false alarms low. This makes AI tools in hospitals safer and more trustworthy for finding diseases like cancer.
What this means in practice
- •For hospital data teams: Detect and flag manipulated medical images during AI diagnosis workflows to reduce missed or false cancer detections caused by adversarial attacks.
- •For medical device engineers: Integrate noise-based adversarial attack detection into AI imaging software to improve clinical reliability and safety certifications.
Authors
Tristan Malatyński, Joanna Jaworek-Korjakowska
Abstract
Deep learning is now central to radiology, helping detect changes on X-rays, CTs, and MRIs. However, these systems are highly vulnerable to adversarial attacks - small, crafted perturbations that mislead models while appearing unchanged to humans. Such errors risk missed cancers or false positives. Studies show over 90% attack success on medical scans. Almost 30% of US hospitals use AI in imaging in 2023. Thus, an effective method of detecting such attacks and their directionality is crucial to ensure the credibility and reliability of medical systems based on DNNs. We propose a novel detection framework that leverages subsequent attacks using random noise to identify adversarial perturbations. Our approach uses the analysis of variations in the Sigmoid function to distinguish genuine medical images from adversarially manipulated ones. Specifically, we compare prediction differences between clean and attacked images, as well as between different adversarially altered versions, to improve detection accuracy. We evaluated our method on three popular medical datasets including 'Chest X-Ray Images for Classification', 'Retinal Fundus Multi-disease Image Dataset' and 'Brain Tumor Dataset' under various attack scenarios, including random noise. Our framework achieved up to 99.8% ACC in identifying adversarial directional attacks, significantly outperforming existing defense mechanisms. It consistently identified adversarial samples across varying attack strengths while keeping false positives low, showing strong reliability and potential for clinical use. Furthermore, our proposal highlights its potential for real-world deployment.