Blockchain-Enabled Multi-Key Fully Homomorphic Encryption for Privacy-Preserving Federated Learning with Counterfactual Explanation in Medical Image Analysis

Le résumé fourni par la source

Blockchain-based federated learning with homomorphic encryption in the medical domain ensures data privacy and security during transmission and aggregation. However, existing approaches face two major challenges: 1. Homomorphic encryption schemes used to encrypt local gradient updates sometimes rely on a single shared key, meaning one compromised key exposes every participant, and 2. There is a lack of trust of the global model since each local model is encrypted, and it is difficult to verify the authenticity of the model. We tackle both issues simultaneously by proposing Blockchain-Enabled Multi-Key Fully Homomorphic Encryption for Privacy-Preserving Federated Learning with Counterfactual Explanation. Our approach lets. 1. Each hospital generates and retains its own cryptographic keys with key-switching protocols for secure cross-key aggregation and threshold decryption. 2. We deploy an Ethereum smart contract that logs each participant's contribution. 3. We propose enhanced counterfactual explanations to explain the global model's decisions. Our counterfactual method outperforms existing approaches with 23% higher validity and 18% better proximity scores. Our Multi-Key Fully Homomorphic Encryption provides independent keys with a moderate overhead increase.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.