Research And Implementation Of Federated Learning Platform Based On Blockchain

The development of artificial intelligence brings great convenience to people,but the formation of data silos hinders its development as people pay more and more attention to data privacy protection.FL(Federated learning),proposed by Google in 2016,is an important technique to solve the problem of data silos.It can train models without revealing participants' datasets.However,traditional FL algorithms still face three security threats.Firstly,malicious nodes may perform data poisoning attacks and contaminate the output models.Secondly,centralized stored data may be tampered which leads to mutual distrust among nodes.Thirdly,malicious nodes may reason about the gradient and obtain some private information.To mitigate data poisoning on FL,this paper proposes VB-Fed AVG(Voting Based Federated Average).VB-Fed AVG introduces the voting strategy in ensemble learning and recalculates the weights of model aggregation.In sybils attack,this paper introduces similarity judgment and proposes its optimized version named VSB-Fed AVG(Voting and Similarity Based Federated Average)to mitigate data poisoning from repeated malicious nodes.The comparison experiments in this paper use three datasets includes MNIST,KDDCUP99 and AMAZON,and six attack scenarios.The result demonstrates that VB-Fed AVG has higher robustness than traditional FL algorithms in a wide range of cases while VSB-Fed AVG has higher robustness in the case of sybils attack.To address the mutual distrust and privacy issues in the FL,this paper proposes ISC-FL(IPFS and Smart Contract Federated learner).ISC-FL introduces ethereum smart contracts to solve the trust problem in FL and utilizes IPFS to optimize the huge overhead and costs in blockchain.Since the transparency of blockchain can aggravate the privacy leakage,this paper designs a multi-key encryption strategy to guarantee the privacy of on-chain data.Additionally,ISC-FL implements VB-Fed AVG to get higher robustness.In this paper,we demonstrate that ISC-FL solves trust and privacy issues through security analysis.Also,this paper conducts timeconsuming analysis and cost analysis through simulation experiments.The result shows that ISCFL achieves a balance between trust,privacy protection and performance.Finally,this paper designs and develops a secure federated learning platform.The system utilizes the ISC-FL framework for federated learning which comprehensively resists security threats to federated learning.Also,the system implements a Web management interface using HTML,CSS,React JS,and Node JS.The system allows users to publish federated learning tasks and participate in federated learning tasks in a highly robust,trusted and privacy-preserving environment.With secure federated learning platform,users can share the results of collaborative training and visually monitor the task process.The implementation of the system further demonstrates the application value of the solution.