MuCALD-SplitFed: Causal-Latent Diffusion for Privacy-Preserving Multi-Task Split-Federated Medical Image Segmentation

Federated learning (FL) allows hospitals to train AI models collaboratively without sharing patient data, but real-world clinical workflows often involve multiple distinct tasks—standard FL and split-FL aren't designed for this. Enter MuCALD-SplitFed, a novel multi-task SplitFed framework from researchers at Simon Fraser University and the University of British Columbia. It combines causal representation learning with a latent diffusion model to handle heterogeneous tasks across decentralized clients while preserving privacy. The system splits the model between clients and a server, reducing client computation, and uses causal constraints to ensure learned features are task-relevant and disentangled, preventing cross-task interference.

Experiments on medical image segmentation datasets show that MuCALD-SplitFed not only converges where baseline SplitFed fails, but also achieves higher Dice scores across diverse tasks. Crucially, it significantly reduces vulnerability to reconstruction and membership inference attacks at the split point, addressing a major privacy concern in distributed learning. The method outperforms state-of-the-art personalized FL and multi-task FL approaches, all while maintaining a small communication overhead. Accepted as an oral presentation at ICIP 2026, this work paves the way for secure, multi-task collaborations among medical institutions with differing imaging specialties.