Application And Key Issues Of Blockchain In Federated Learning

Rice is a major cash crop in China and its fine management contributes to food security and the stable and sustainable development of agriculture.Federated learning is suitable for solving the problems of Data silos,the low utilization of information,and the asymmetry of information in the intelligent refinement of rice management.Federated learning shares training results without data transmission and can largely protect the privacy of original data,but still exists problems such as single-point failure,model inference attacks,and malicious participants.This paper investigates machine learning algorithms incorporating adaptive differential privacy protection to address model reasoning attacks.It explores strategies to counteract poisoning attacks in federated learning models by leveraging adaptive differential privacy protection.Additionally,the study proposes verifiable federated learning solutions that integrate blockchain technology and differential privacy for secure and reliable data sharing in prototype system for rice leaf pest and disease identification The main innovative contributions of this research are outlined below:1.This study introduces a machine learning algorithm that incorporates adaptive differential privacy protection,focusing on mitigating model inference attacks.To counteract such attacks,the algorithm employs truncated Centralized Differential Privacy(tCDP)to perturb the gradient during the local model training process of participants.Furthermore,it achieves adaptive allocation of privacy budget through adaptive step size and dynamic privacy budget adjustment.The algorithm is evaluated against four baseline differential privacy protection algorithms using publicly available datasets.Experimental and theoretical analyses demonstrate that the proposed algorithm effectively balances data privacy and model learning performance,even under the constraint of a fixed privacy budget.2.This study proposes a verifiable federated learning scheme based on blockchain and differential privacy,building upon the federated learning model with adaptive differential privacy protection.To ensure the data privacy of all participants in the federated learning process,adaptive differential privacy protection is introduced.Furthermore,to address the issues of single point of failure and potential malicious behavior by the server,the central server in the federated learning model is replaced by a decentralized blockchain network.To counteract malicious behavior by participants,a participant committee is established using smart contracts.Additionally,Pearson similarity is utilized to measure model feature differences,and both of these blockchain verification methods are employed to mitigate the impact of parameters from malicious nodes.Experimental results and security analysis demonstrate the effectiveness of this scheme in eliminating the influence of malicious nodes.Moreover,the scheme remains robust against malicious parties even when they exceed half of the total participants.3.Based on the verifiable federated learning scheme which is based on blockchain and differential privacy,a prototype system for rice leaf pest and disease identification with safe and reliable data sharing is realized.This system involves the three stages of rice planting,that is,pre-production,production,and post-production,and supports various users such as growers,purchasers,and individual users,in which the federated learning is used to identify rice leaf disease,and the accuracy can reach 90%.The final test results show that the prototype system can provide secure data sharing and information flow while performing the expected functions well.