Research On Key Technologies On Federated Learning Based On Blockchain

Federated learning is a machine learning framework with data privacy protection capabilities.In federated learning,the system cannot completely control the behavior of clients,and clients may have abnormal behaviors such as uploading error gradients or delayed uploading due to Byzantine attacks or its own software and hardware failures.In federated learning,the parameter server may also have abnormal behavior in the process of gradient collection and aggregation,mainly including gradient-based client raw data inference and model parameter incorrect update.The above unresolved issues may lead to reduced accuracy of models trained by federated learning frameworks,leakage of user data privacy,etc.Existing research has improved and enhanced the robustness of federated learning by utilizing the decentralization and immutability of blockchain.For untrusted clients,most studies are based on Byzantine Fault Tolerance,which is untargeted defense against clients,and there is a problem that the accuracy of the model decreases.In addition,most of the research work is on unencrypted gradients,and there is insufficient research on dealing with client-side anomalies in the case of gradient encryption.For untrusted parameter servers,existing research work has problems in resource consumption and scalability.Aiming at the problems of existing study,this paper studies the robustness of federated learning,and proposes a parameter update security concerned federated learning architecture based on blockchain.The main work of the paper includes four parts:(1)A parameter update security concerned federated learning architecture based on blockchain is proposed.The auditability of the parameter iteration process is realized by using the immutability of the blockchain.A reputation scoring model for federated learning clients is proposed,and the anomaly detection algorithm is used to effectively filter malicious parameters.Combined with the incentive mechanism of the blockchain,the fair measurement of clients’ contribution and the effective punishment of abnormal behavior are realized.(2)A parameter anomaly detection algorithm based on homomorphic encryption is proposed.Through the threshold encryption algorithm,the independent encryption and cooperative decryption of the client parameters are realized,and the KNN-based anomaly detection algorithm is designed for the reputation score of the parameters,which improves the accuracy of the model and makes it difficult to implement the original data inference attack on the parameters.(3)A parameter aggregation mechanism with enhanced scalability is proposed.Through the replacement of consensus organization based on VRF and the election of leader nodes based on reputation,the nodes with participation rights in the consensus process are maintained in a small scale with high reliability range,which reduces the communication complexity of the consensus algorithm.It solves the single point of failure problem of the parameter server while maintaining scalability.(4)The feasibility and effectiveness of the above architecture and algorithm are proved by experiments.Aiming at the task of image classification on the FEMNIST dataset,the paper conducts experiments on the proposed algorithm and existing schemes,and compares and analyzes the experimental results.In the comparative experiments,it is verified that the parameter anomaly detection algorithm proposed in this paper has a performance improvement of 5%-27%in accuracy.Experiments verify that the scalability-enhanced parameter aggregation consensus algorithm proposed in this paper reduces the communication complexity from a linear level to a constant level in a small-scale network.