Leveraging data drift to optimize differential privacy in federated learning via dual-mode switching

Federated Learning (FL) is a distributed machine learning paradigm that enables multiple clients to collaboratively train a shared model while keeping their data local. In real-world deployments, FL remains vulnerable to privacy attacks that may lead to the leakage of sensitive information. Differential Privacy (DP) is widely adopted in FL to mitigate such risks, but often at the cost of degraded model performance. However, existing DP methods optimize the privacy-utility trade-off by only considering the direct impact of noise on model performance, overlooking the significant data drift that naturally occurs during distributed training in FL. Crucially, this data drift inherently degrades model performance, and the additional DP noise further exacerbates this degradation, severely impairing model convergence and accuracy. To address this issue, we propose a novel DP framework named Dual-mode Switching Differential Privacy (DSDP), which leverages data drift to improve model performance while preserving strong privacy guarantees. First, to monitor the degree of data drift, we design two evaluation metrics: gradient sensitivity and client contribution, which together form a dual-dimensional evaluation strategy. Second, based on this evaluation strategy, we implement a dynamic privacy budget allocation mechanism. Clients that contribute positively to model training are assigned higher privacy budgets, allowing reduced noise and improved accuracy of the global model. Third, to directly counter the convergence slowdown exacerbated by data drift, we propose a dual-mode noise switching mechanism: Gaussian noise is applied to accelerate convergence during the early stages of training, while Laplace noise is used to maintain strong privacy guarantees as training progresses. Extensive experiments on four benchmark non-IID datasets demonstrate that DSDP outperforms state-of-the-art DP methods including DP-FedProx, PrivateFL, and NbAFL in terms of model accuracy. Moreover, under model poisoning and gradient leakage attacks, DSDP achieves a superior balance between privacy and utility while exhibiting strong robustness.