Evaluating vulnerabilities and protections of AI methods in computer-assisted image-based diagnosis

Abstract

This paper examines three safety aspects of AI in medical image analysis: vulnerability to adversarial attacks, patient data protection via generative models, and federated learning without data sharing. Experiments on over 260,000 chest X-rays, CT slices, and histological slides show that white-box adversarial attacks (FGSM, AutoAttack, Carlini-Wagner) can reduce classification accuracy to 0% when no defense is applied. Among three defense strategies, a high-level task-driven denoiser proved most effective, restoring accuracy up to 100%. Generative models (DC-GAN, Pro-GAN, diffusion) produced realistic synthetic images; using them for data augmentation led to only a modest accuracy drop (2.2-3.5% for deep learning). Federated learning (FedAVG) succeeded only for homogeneous datasets (chest X-rays) but failed for highly variable histological images. The paper concludes that adversarial attacks pose critical threats, generative models enable privacy-preserving augmentation, and federated learning requires careful modality-specific adaptation.