Investigating privacy leakage in distributed deep learning for medical imaging through gradient inversion attacks

Federated learning has gained considerable attention recently in medical image analysis as it enables model training without data sharing. On the other hand, studies show that exchanged gradients can reveal information through gradient inversion attacks which raise concerns about privacy in application areas such as healthcare. Results are available to observe leakages and suggest analysis of deep learning models on medical images but there exists the need for studying the impact gradient leakage by different models. This study investigates privacy leakage in federated learning using the NIH Chest X-ray dataset by evaluating the susceptibility of three deep learning architectures namely Simple CNN, ResNet-50, and EfficientNet to gradient-based image reconstruction attacks. In this work, privacy leakage is assessed using heatmaps and reconstruction overlays. Experimental results show that EfficientNet consistently achieves higher reconstruction fidelity than the other evaluated models. Our work suggests that model architecture can deeply influence the extent of information that can be recovered from shared gradients. We also analyze privacy risks in medical imaging and highlight the importance of incorporating stronger privacy-preserving mechanisms when deploying federated learning in healthcare domain.