Data-Aware Layer Assignment for Secure and Efficient Communication in Federated Learning for Medical Image Analysis

Abstract

Cross-silo medical imaging federations must contend with strict privacy, limited bandwidth, and non identically distributed (non-IID) data that destabilize training. Current federated learning (FL) architectures either carry the full model (e.g., FedAvg/FedProx) or use naive client/layer pruning and random sampling while ignoring both non-IID heterogeneity and per-layer utility. Based on these limitations, the paper presents a dataaware, layer-wise protocol that aligns communication with expected loss descent while bounding per-round client leverage. Each round, the server estimates per-layer influence from a tiny root set, and clients expose lightweight metadata to form data-quality scores. A capacity-constrained entropic transport matches high-influence layers to high-quality clients under redundancy and temporal coverage. Clients train all layers but upload exactly one with train-all, send-one principle. The server then performs per-layer robust aggregation on masked updates via secure aggregation. On the three cross-silo imaging benchmarks of Pneumonia CXR, Brain-Tumor MRI, and ISIC Skin Cancer, it demonstrates a strong threshold free detection quality (AUROC/ AUPRC: 0.925/0.935, 0.996/0.988, 0.834/0.852, respectively) while also reducing the per round up-link by ≈ 1/n with respect to FedAvg (e.g., ≈ 10× with 10 clients) by only receiving one layer per client. Indicating its viability for deployment-grade secure aggregation for hospital networks.