Sppdp-fl: adaptive client selection and dimensional projection for performance enhancement in personalized differential privacy federated learning

Federated learning is a privacy-preserving distributed machine learning framework that trains models over a number of clients while keeping data localized. To better mitigate the risks associated with information leakage, one common technique is adding noise to the model or gradient before uploading those parameters to the server. Different clients might have their own preferences regarding data privacy requirements, which leads to personalized differential privacy federated learning scenarios. Personalized privacy budget leads to different noise interference toward client models, and the noise interference affects the central server’s model aggregation, which deteriorates the global model’s performance. This paper proposes a novel personalized differential privacy federated learning scheme that makes significant advancements in two key areas. First, a dynamic client selection strategy is introduced, which addresses the challenge of balancing privacy and model performance. By dynamically adjusting the selection probability based on the clients’ decreasing privacy budgets, this strategy optimizes the contribution of clients with higher privacy budgets, improving model efficiency and personalization over the traditional random selection method. Second, an enhanced differentially-private stochastic gradient descent algorithm is employed, which reduces the adverse impact of noise on model accuracy. By projecting noisy gradients into a low-dimensional subspace using the top-k eigenvectors of the second-order moment matrix, the algorithm effectively preserves critical information while maintaining privacy. Experimental results demonstrate that the proposed scheme achieves a 10% improvement in accuracy compared to baseline methods in the IID and non-IID environments.