A Novel Differential Privacy Federated Learning Framework: An Adaptive Budget Allocation and Reversion Method

In computer multimedia, Federated Learning (FL) enables clients to contribute their multimedia data, thereby improving the overall accuracy of the model. However, clients’ sensitive multimedia data could still be inferred through the submitted local models in FL. As a result, differential privacy (DP) techniques have been adopted to protect clients’ sensitive multimedia data. But existing works primarily focused on adding fixed noise to data, gradients, or loss functions, which can negatively affect the model’s accuracy and convergence. To address this issue, we propose a novel differential privacy federated learning framework (DP-FedAR) which consists of two modules, i.e., an adaptive budget allocation method and a reversion mechanism. Specifically, based on model similarity, an adaptive allocation rule is proposed to assign privacy budgets in real time for each training round. Then, in order to avoid exhausting the privacy budget of each client too early in the whole training process, a reversion mechanism is further devised to identify clients’ historical models that mostly resemble the current global model. Theoretical analyses demonstrate that our proposed DP-FedAR can converge and has a strict privacy guarantee. Moreover, extensive simulations validate that our proposed DP-FedAR outperforms existing algorithms in terms of training accuracy. More precisely, the accuracy of DP-FedAR surpasses that of counterparts, with an average improvement of 8% to 12%