Blockchain-based Federated Learning For Secure Medical Data Modeling

The development of technology has led to a large amount of medical data stored in medical institutions,which have become data islands due to the consideration of medical privacy laws and regulations,resulting in an unbalanced distribution of medical resources.The emergence of federated learning is considered a good method for medical data processing,where data can be stored locally while only the trained parameters are uploaded,which can prevent the leakage of raw medical data and well address the concerns of patients and medical institutions about medical privacy leakage.However,the traditional way of centralized server storage is prone to single point of failure.Meanwhile,considering the self-interest and mutual distrust among medical institutions,this paper proposes a distributed medical data modeling and sharing framework combining two emerging technologies,federated learning and blockchain,to solve the above problems.Blockchain,as the foundation of federated learning,helps to provide a distributed,secure,transparent,and traceable learning and sharing environment.At the same time,the 51%unattackability of blockchain makes the data stored in it very secure that An attacker cannot successfully corrupt all the data stored in the blockchain by attacking a single block,making the single point of failure problem a thing of the past.The main work of this paper is as follows:(1).In the blockchain-based system,this paper uses homomorphic encryption and writes attribute-based access control policies into smart contracts,which can well solve the security risks of medical models during uploading and storage through medical model encryption and decryption,identity verification and assignment and update of identity rights to medical institutions.(2)This paper proposes a contribution-weighted federated learning aggregation algorithm and its associated contribution-weighted incentive mechanism,which fully considers and guarantees participants’ contributions and corresponding rewards,prompting participants to more actively provide more medical data to participate in training,and solving the trouble of difficult medical data collection.(3)Finally,this paper builds a Hyperledger Fabric blockchain network environment and implements the validation by simulating several medical institution nodes with several publicly available medical datasets.The results show that the blockchain-based contribution-weighted aggregated federated learning outperforms both centralized machine learning methods and traditional federation learning average aggregation methods in terms of model accuracy and system security.