Research On Data Privacy Protection Method Of Federated Learning Across Data Silos

The privacy problem restricts the use and sharing of data,which makes all parties with data form silos of data.The combination of federated learning and homomorphic encryption or differential privacy can break this limitation and achieve federated learning across silos.Federated learning based on homomorphic encryption can protect user data privacy while implementing distributed machine learning,but the encrypted data is transparent and cannot be detected for correctness,so it is not sufficiently tolerant to malicious nodes.The use of differential privacy can protect the data privacy of each node during federated learning,but adding noise to the model parameters will reduce the accuracy and convergence efficiency of the model.In order to address the problem of insufficient fault tolerance,this paper improves the federated learning and proposes the Fault-tolerant Homomorphic Encryption Federated Learning Model Based on Grouping Aggregation(GFTFL).GFTFL is aggregated twice in each iteration.The first aggregation divides the participating nodes into several contribution groups,and encrypts the training results of each contribution group using a homomorphic encryption scheme based on the Paillier algorithm,and decrypts the aggregation results of the corresponding contribution group only after aggregating each contributor group to achieve gradient privacy protection.The second aggregation uses the global test set to test the training results of each contributor group,and selects a number of contributor groups with higher accuracy to participate in the aggregation to obtain the final training results to achieve the fault tolerance mechanism.GFTFL achieves user gradient privacy protection and local model detectability in the presence of malicious nodes environment,and then selects a portion of contributor groups with better test results to participate in the aggregation to improve the federal learning process for The ability of the federation learning process to tolerate malicious nodes.The experimental results show that GFTFL is more fault-tolerant than the traditional scheme in the presence of malicious nodes.To address the problem that noise affects the correct rate,this paper proposes the Kalman Filter-based Differential Privacy Federated Learning Model(KDPFL),which converts the model parameter update equation into a state transfer difference equation,and calculates the Kalman gain based on the state update difference equation with The Kalman gain is calculated based on the state update differential equation and the noise distribution,and the Kalman gain coefficient is used to correct the parameter update to achieve the noise reduction.And the validation scheme of effectiveness is constructed in the case of two datasets,non-independent identically distributed(Non-IID)dataset and independent and identically distributed(IID)dataset,and the experiment results show that the accuracy of the proposed method in this paper is improved by about 3% compared with differential privacy federated learning.