Scholarlink

<p>Recent advancements in artificial intelligence have heightened concerns about data privacy. While differential privacy is commonly used in federated learning to protect model parameters, a significant challenge remains: balancing user privacy with model performance. This paper proposes a novel approach, the Federated Learning Method Based on Adaptive Differential Privacy and Homomorphic Encryption (ADPHE-FL), to address these challenges. The proposed scheme dynamically adjusts the privacy budget across communication rounds, offering a more flexible and efficient privacy management mechanism. Additionally, a mean-k parameter clipping algorithm is introduced to reduce transmitted data efficiently, optimizing model training and significantly enhancing both communication efficiency and system performance. A two-layer privacy protection strategy is employed to ensure secure aggregation while mitigating issues related to data utility degradation and high computational costs. Experimental results demonstrate that, with a privacy budget of 1, ADPHE-FL outperforms existing methods across multiple datasets, achieving a 2% accuracy improvement over CLFLDP and 18% over DP-FL on the CIFAR-10 dataset at convergence. Furthermore, with a key length of 1024, encryption time efficiency improves by 89.3%, and decryption time efficiency by 82.6%, compared to the traditional Homomorphic Encryption-based FedAvg method.</p>

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

ADPHE-FL: Federated learning method based on adaptive differential privacy and homomorphic encryption

  • 期刊论文

  • 发表日期：

    2025年4月5日

查看全文
  • Tao Wu,
  • Yulin Deng,
  • Qizhao Zhou,
  • Xi Chen,
  • Ming Zhang

摘要

Recent advancements in artificial intelligence have heightened concerns about data privacy. While differential privacy is commonly used in federated learning to protect model parameters, a significant challenge remains: balancing user privacy with model performance. This paper proposes a novel approach, the Federated Learning Method Based on Adaptive Differential Privacy and Homomorphic Encryption (ADPHE-FL), to address these challenges. The proposed scheme dynamically adjusts the privacy budget across communication rounds, offering a more flexible and efficient privacy management mechanism. Additionally, a mean-k parameter clipping algorithm is introduced to reduce transmitted data efficiently, optimizing model training and significantly enhancing both communication efficiency and system performance. A two-layer privacy protection strategy is employed to ensure secure aggregation while mitigating issues related to data utility degradation and high computational costs. Experimental results demonstrate that, with a privacy budget of 1, ADPHE-FL outperforms existing methods across multiple datasets, achieving a 2% accuracy improvement over CLFLDP and 18% over DP-FL on the CIFAR-10 dataset at convergence. Furthermore, with a key length of 1024, encryption time efficiency improves by 89.3%, and decryption time efficiency by 82.6%, compared to the traditional Homomorphic Encryption-based FedAvg method.