Scholarlink

Federated learning (FL) is currently popular distributed machine learning technology, which enables the iterative optimization of global models while safeguarding user data privacy. However, users’ private information may still be inferred through analysis of model parameters after training, such as the weights in deep neural networks. Many methods currently exist to protect model parameters using local differential privacy (LDP) in federated learning, but they struggle to achieve an optimal balance between privacy protection and model accuracy. This paper introduces the Positive and Negative Piecewise Mechanism (PNPM), which introduces perturbations to local model parameters prior to aggregation. First, it is demonstrated that the mechanism satisfies the strict definition of differential privacy, ensuring the privacy of the algorithm. Second, it is shown that the mechanism can maintain model accuracy even with fewer users, ensuring its effectiveness. Finally, compared with other state-of-the-art methods in terms of model accuracy and privacy preservation, PNPM demonstrates superior performance, achieving a global model test accuracy deviation of less than 10% on the MNIST, Fashion-MNIST, and CIFAR-10 datasets when the privacy budget \( \epsilon \) is set to 1.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An Efficient Federated Learning Privacy Preservation Method with Differential Privacy Against Model Inversion Attack

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Bolun Wang,
  • Dong Wang,
  • Chenpu Li,
  • Jinhuan Wang

摘要

Federated learning (FL) is currently popular distributed machine learning technology, which enables the iterative optimization of global models while safeguarding user data privacy. However, users’ private information may still be inferred through analysis of model parameters after training, such as the weights in deep neural networks. Many methods currently exist to protect model parameters using local differential privacy (LDP) in federated learning, but they struggle to achieve an optimal balance between privacy protection and model accuracy. This paper introduces the Positive and Negative Piecewise Mechanism (PNPM), which introduces perturbations to local model parameters prior to aggregation. First, it is demonstrated that the mechanism satisfies the strict definition of differential privacy, ensuring the privacy of the algorithm. Second, it is shown that the mechanism can maintain model accuracy even with fewer users, ensuring its effectiveness. Finally, compared with other state-of-the-art methods in terms of model accuracy and privacy preservation, PNPM demonstrates superior performance, achieving a global model test accuracy deviation of less than 10% on the MNIST, Fashion-MNIST, and CIFAR-10 datasets when the privacy budget \( \epsilon \) is set to 1.