The Hidden Adversarial Vulnerabilities of Medical Federated Learning

TL;DR

This paper reveals new vulnerabilities in federated medical image analysis systems, showing how adversaries can exploit gradient information to improve attack efficiency and transferability, highlighting security concerns in healthcare AI.

Contribution

It uncovers a novel attack method leveraging gradient info from global models, demonstrating that single-step attacks can be more efficient than iterative ones in federated medical settings.

Findings

Gradient-based attacks are more effective with prior global model info.

Single-step attacks outperform iterative ones in efficiency.

Highlights security risks in federated healthcare AI.

Abstract

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.