Secure Federated Learning In Healthcare Using Blockchain And Smpc

Abstract

In the World of AI in Healthcare, coupling Federated Learning (FL) with Blockchain is a changing paradigm in building secure and privacy upholding AI systems. However, FL potentially can be subject to model poisoning attacks and has no mechanisms for integrity verification of local models. This paper presents a blockchain-based federated learning framework with Secure Multi-Party Computation for the verification of encrypted models. Prior to aggregation, local model validation takes place under privacy preservation to ensure no malicious updates are counted. Verified models are then stored and SMPC aggregation on the blockchain to enable tamperproof decentralized training. The updated global model is shared among participants via the blockchain ledger. Experimental evaluations using Convolutional Neural Networks (CNNs) on medical datasets show that the proposed system can eliminate all poisoned models, improving global model accuracy from 0 to potentially 25%, while the verification speed is still close to normal inference. This framework promotes trust, data privacy, and model integrity in collaborative healthcare AI.