Adaptive differential privacy federated learning with client self selection

With the wide application of AI services in the era of big data, federated learning models that can solve the data silo problem without sharing sensitive data in different devices have received much attention. Research in recent years has shown that private information can still be inferred by analyzing model parameters localized in federated learning models. To address this risk, differential privacy techniques have been applied to federated learning models due to their unique privacy-preserving approach to protect customers’ private data. However, the proposed joint learning models based on differential privacy still have flaws. Firstly, the heterogeneity of the gradient parameters is not taken into account which affects the convergence of the model and the quality of the training parameters, because the gradient cropping thresholds used in the model training are the same, and it is not possible to adaptively adjust the amount of added noise. Secondly the way of client selection in the model is a random selection method, which is not conducive to ensuring that excellent clients are selected to participate in model aggregation in each round of training. Based on the above problems, this paper proposes a federated learning model with aggregated gradient adaptive cropping technique and client self-selection, which adapts to adjust the amount of noise by means of adaptive gradient cropping of different clients in different rounds, and combines roulette and elite retention client sampling methods to accelerate the convergence of the model. Experiments demonstrate that our proposed model is able to improve the classification accuracy of the final model by 5.1% under the same level of privacy constraints compared to the traditional joint learning model. In terms of convergence speed, the number of rounds required for our model to enter the convergence state is reduced by $\mathbf{1 0 - 1 5}$ rounds compared to the traditional approach.