Research On Privacy Protection Algorithm Of Federated Learning Based On Alliance Chain Committee Consensus Mechanis

With the proliferation of the amount of data generated by different devices in Io T,there is a need for centralized servers to collect large amounts of data as training sample sets for machine learning models to perform high-quality model training.The reluctance of many users to share local data for model training has led to the existence of "data silos",coupled with the fact that these large amounts of data often contain privacy-sensitive information.To address these two problems,federated learning was developed.While federated learning does not transmit data directly,but rather gradient information,it allows participating users to train models together using multiple parties’ data without revealing local data,making the data available but not visible.It has been demonstrated that gradients can potentially leak private information.Within the current federated learning environment,user data may encounter privacy leakage issues such as malicious nodes providing low-quality model parameters to corrupt the global model during training,resulting in a negative impact on the training outcome.Additionally,there are risks of attacks from malicious servers or inference attacks,as well as data tampering,all of which can lead to a lack of trust among nodes.To address these issues,this paper layers a series of federated learning-based approaches to address these problems.The main three approaches are as follows: 1)Firstly,a differential privacy-based federated learning privacy-preserving scheme DPFL is proposed,which mainly achieves the protection of local model parameters through Gaussian mechanism in the FL process.2)Adopting blockchain instead of centralized server in FL system avoids malicious centralized server to obtain data privacy using inference attack to ensure users’ collaborative training and achieve decentralization.Meanwhile,the use of proof-of-work mechanisms to validate and broadcast model updates,coupled with the fact that data submitted to the blockchain cannot be tampered with,helps federated learning track down the root cause of problems when they arise during training.3)Leveraging consortium chains instead of the centralized server in traditional federated learning to verify local model updates as well as global model generation and storage.Adding perturbations during local training using Gaussian mechanism for initial privacy preservation.The model parameters are then validated by committee nodes to reduce the consensus cost,and the receipt of sufficient validated model parameters in the update block will trigger a smart contract to aggregate the updates and broadcast the new global model to the nodes for the next round of training.Coupled with the incentive mechanism based on the model accuracy scores,more honest and reliable users are encouraged to participate in the training process of the model.Finally,this paper adopts a convolutional neural network,based on federated learning,to experiment with the MNIST dataset and use the MNIST dataset to test the performance of the proposed model.In this paper,we compare the corresponding metrics of model performance,consensus cost,differential privacy budget,and malicious node tolerance ratio in Basic FL,DPFL,BFL,and CFLPP frameworks,respectively.Overall,the results show that the CFLPP framework outperforms the other three frameworks in terms of model performance for a fixed malicious nodes proportion and outperforms the DPFL framework and the BFL framework in terms of resistance to malicious node attacks due to the use of committee nodes for verification of model updates,which greatly reduces the consensus time compared to the BFL framework.Additionally,the experiments also demonstrate that a lower privacy budget provides more robust privacy protection.Therefore,the experimental results demonstrate the feasibility,security,and effectiveness of the CFLPP framework.