A Blockchain-Federated Learning Framework for Privacy-Preserving Medical Image Analysis with Adaptive Differential Privacy and Zero-Knowledge Proof Aggregation

Abstract

Abstract Federated learning with blockchain technology is a promising domains to facilitate privacy-sensitive collaborative intelligence in the domain of medical imaging. Traditional federated learning systems are affected by several classes of vulnerabilities, including dependence on a single aggregator, susceptibility to gradient-inversion attacks, Byzantine attacks and model poisoning, and the lack of a tamper-proof audit mechanism required by different healthcare standards. The paper presents a blockchain-based Federated Learning framework for addressing these limitations by proposing an Adaptive Differential Privacy mechanism (ADP) that dynamically calibrates the noise scale σt across federated rounds. For the evaluated configuration (σ0 = 1.5, T = 100 rounds), this schedule composes under standard R´enyi-DP accounting to a cumulative budget of ε ≈ 244 at δ = 10−5. A Zero-Knowledge Proof Aggre- gation Protocol (ZK-AP) enables verifiable on-chain model aggregation without disclosing individual participant gradients in plaintext to any party below a decryption-committee threshold, certifying the integrity of the encrypted-submission pipeline rather than the genuine execution of local neural-network training, and a Reputation-Weighted Byzantine FaultTolerance (RW-BFT) consensus mechanism supports heterogeneous medical FL networks. We evaluated our framework on the BraTS 2020, CheXpert, and ISIC 2019 datasets dis- tributed over ten simulated hospital nodes with realistic non-IID data partitions, achieving accuracy within 0.8 percentage points of a centralized Oracle benchmark.