Federated learning under differential privacy with effective clipping and analytic Gaussian mechanism
Jianhui Zhang, Nianming Xue, Xiangyang Y. Li, Youqun Long, Entang Li, Chonghao Xu, Tao Liu, Fangfang Chen · 2025
The utilization of convolutional neural networks has become pervasive across numerous sectors, including areas like visual identification and linguistic processing. Yet, there is a growing concern among users regarding the vulnerability of their privacy, stemming from the accessibility of model parameters. Then we take inspiration from differential privacy. By adding Gaussian noise to the training parameters, we finally realize the privacy preservation. This study implements an optimized image recognition technique that utilizes gradient layer clipping to enhance the efficacy of the model while maintaining differential privacy. Additionally, we employ the analytical Gaussian mechanism to bolster our approach. Findings indicate that at a privacy parameter ϵ of 8.0, the model achieves an accuracy rate of 96.46% on the MNIST dataset and 61.10% on the CIFAR-10 dataset; Meanwhile, with a privacy parameter ϵ of 2.0, the model's accuracy rate is 94.0% for MNIST and 58.79% for CIFAR-10.