ChainLearn: A Blockchain-Based Capacity-Aware Framework for Federated Ensemble Learning

2026-05-23

AI Summary

This work addresses the challenge of heterogeneous computational resources across hospitals in medical imaging federated learning, which often renders standard algorithms ineffective. The authors propose a heterogeneous federated ensemble learning framework that does not require a unified model architecture. By evaluating each participant’s computational throughput, the system dynamically assigns tailored models—such as MobileNetV3-Small, EfficientNet-B0, or ResNet-50—and performs weighted ensemble inference off-chain. A blockchain-based mechanism, implemented via Solidity smart contracts, manages participant registration, performance metrics, and ensemble weights, while only uploading hashes and scalar values to preserve privacy and efficiency. Experiments on PneumoniaMNIST and DermaMNIST demonstrate that the method achieves calibration error no worse than uniform-weight ensembles and accuracy comparable to FedAvg, with a per-round communication overhead of merely 224 bytes—over 910,000 times lower than FedAvg.

Abstract

Federated learning is used in medical imaging where privacy prohibits centralizing data. Standard federated algorithms assume homogeneous hardware, identical architectures, and centralized aggregation, which fails when hospitals have unequal compute resources. We propose capacity-aware coordination: measure each hospital's throughput, assign capacity-appropriate architectures (MobileNetV3-Small, EfficientNet-B0, ResNet-50), and combine predictions via weighted ensemble. Weak and strong hospitals can participate without forcing uniform architectures. We separate on-chain policy from off-chain learning. A Solidity contract stores hospital registration, benchmark hashes, metrics, and weights. Hospitals train locally and submit only hashes and scalars (not parameters). Weighted ensemble inference is computed off-chain. Experiments on PneumoniaMNIST and DermaMNIST (5 seeds, 3 non-IID levels) show our method achieves lower or equal calibration error versus equal-weight ensemble and competitive accuracy versus FedAvg, FedProx, and FedMD. Communication overhead is 224 bytes per round, a reduction of over 912,000x compared to FedAvg.

Problem

• federated learning
• heterogeneous hardware
• medical imaging
• model heterogeneity
• resource inequality

Innovation

• capacity-aware federated learning
• blockchain-based coordination
• heterogeneous model ensemble
• communication-efficient FL
• on-chain off-chain separation