Efficient adaptive defense scheme for differential privacy in federated learning

Abstract

Federated learning, as an emerging technology in the field of artificial intelligence, effectively addresses the issue of data islands while ensuring privacy protection. However, studies have shown that by analyzing gradient updates, leaked gradient information can still be used to reconstruct original data, thus inferring private information. In recent years, differential privacy techniques have been widely applied to federated learning to enhance data privacy protection. However, the noise introduced often significantly reduces the learning performance. Previous studies typically employed a fixed gradient clipping strategy with added fixed noise. Although this method offers privacy protection, it remains vulnerable to gradient leakage attacks, and training performance is often subpar. Although subsequent proposals of dynamic differential privacy parameters aim to address the issue of model utility, frequent parameter adjustments lead to reduced efficiency. To solve these issues, this paper proposes an efficient federated learning differential privacy protection framework with noise attenuation and automatic pruning (EADS-DPFL). This framework not only effectively defends against gradient leakage attacks but also significantly improves the training performance of federated learning models.

Extensive experimental results demonstrate that our framework outperforms existing differential privacy federated learning schemes in terms of model accuracy, convergence speed, and resistance to attacks.