Federated Learning (FL) is emerging as a robust approach to distributed machine learning where data privacy and decentralized computation are vital. By training models directly on end-user devices and sharing only model updates with a central server, FL offers inherent privacy benefits over traditional centralized training paradigms. However, even without transferring raw data, model updates such as gradients can inadvertently leak sensitive information, compromising user privacy. This paper investigates the potential for privacy leakage in FL systems and introduces a novel adaptive differential privacy mechanism that adjusts noise levels based on the gradient sensitivity of each participating client. We define gradient sensitivity using the L2 norm of each client's model updates and propose a scalable, lightweight, and effective framework that injects privacy-preserving Gaussian noise proportionally. Our contributions include a formal definition of gradient norm sensitivity, a robust adaptive noise injection technique, and an in-depth empirical evaluation on the MNIST dataset. Comparative analysis with FedAvg and fixed-noise DP baselines shows that our method, GS-ADP (Gradient-Sensitive Adaptive Differential Privacy), achieves strong privacy guarantees with minimal accuracy loss. This research contributes a deployable and efficient solution to improve the privacy-utility tradeoff in federated learning architectures, especially for edge-AI and privacy-critical applications. (McMahan et al., 2017; Bonawitz et al., 2019) (Dwork Abadi et al., 2016) (Nasr et al., 2019; Melis et al., 2019)
Ulisi et al. (Thu,) studied this question.