The Hidden Adversarial Vulnerabilities of Medical Federated Learning
This work highlights a critical security problem for federated healthcare systems, revealing hidden adversarial vulnerabilities that could compromise patient data and model integrity.
The paper tackled the vulnerability of federated medical image analysis systems to adversarial attacks by showing that using gradient information from prior global model updates can enhance attack efficiency and transferability, with single-step attacks outperforming iterative ones in efficiency while reducing computational demand.
In this paper, we delve into the susceptibility of federated medical image analysis systems to adversarial attacks. Our analysis uncovers a novel exploitation avenue: using gradient information from prior global model updates, adversaries can enhance the efficiency and transferability of their attacks. Specifically, we demonstrate that single-step attacks (e.g. FGSM), when aptly initialized, can outperform the efficiency of their iterative counterparts but with reduced computational demand. Our findings underscore the need to revisit our understanding of AI security in federated healthcare settings.