DiffusedSplitFed: Latent Diffusion and global feature fusion meet Split Federated medical image segmentation

TMLR Paper6011 Authors

Abstract: Federated Learning (FL), Split Learning (SL), and Split Federated Learning (SplitFed) are emerging paradigms in privacy-preserving medical image analysis. FL enables multiple clients in collaborative model training without raw data exchange, while SL reduces client-side burden by partitioning the model between client and server. SplitFed combines the strengths of both but often faces limited representation power and semantic loss at the client-server interface, affecting both performance and privacy. The intermediate features and gradients transmitted can still reveal patterns from the original data, making them vulnerable to reconstruction attacks. This poses serious privacy risks, especially in sensitive domains like healthcare. This study proposes \textbf{DiffusedSplitFed}, the first SplitFed framework integrating Latent Denoising Diffusion Models (LDDMs) at both forward and backward split points to obfuscate transmitted representations. We design and compare three architectural variants (V1-V3) that explore dual conditioning and global feature fusion on segmentation performance, privacy preservation, and deployment complexity. We evaluated our framework on multiple medical imaging datasets, demonstrating significant segmentation performance while ensuring privacy and robustness compared to traditional SplitFed, state-of-the-art generative baselines, and privacy resilience baselines. We also provide a theoretical convergence guarantee. Our results underscore the potential of latent diffusion and global fusion for privacy-aware, high-fidelity medical image analysis. The implementation is available at: \url{https://anonymous.4open.science/r/DiffusedSplitFed}.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Jiangchao_Yao1

Submission Number: 6011