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

Researchers introduced MuCALD-SplitFed, a multi-task Split-Federated Learning framework combining causal representation learning with latent diffusion models for privacy-preserving medical image segmentation. The architecture partitions models between clients and servers while embedding causal latent spaces to reduce reconstruction-based privacy leakage at split points. Baseline SplitFed failed to converge on multi-task settings where MuCALD-SplitFed achieved stable segmentation, a critical gap for real clinical deployments. Federated ML engineers and medical imaging teams building multi-institutional pipelines should prioritize this approach over standard FedAvg or vanilla SplitFed configurations.