SPM-FL: A Federated Learning Privacy-Protection Mechanism Based on Local Differential Privacy

Federated learning is a widely applied distributed machine learning method that effectively protects client privacy by sharing and computing model parameters on the server side, thus avoiding the transfer of data to third parties. However, information such as model weights can still be analyzed or attacked, leading to potential privacy breaches. Traditional federated learning methods often disturb models by adding Gaussian or Laplacian noise, but under smaller privacy budgets, the large variance of the noise adversely affects model accuracy. To address this issue, this paper proposes a Symmetric Partition Mechanism (SPM), which probabilistically perturbs the sign of local model weight parameters before model aggregation. This mechanism satisfies strict ϵ-differential privacy, while introducing a variance constraint mechanism that effectively reduces the impact of noise interference on model performance. Compared with traditional methods, SPM generates smaller variance under the same privacy budget, thereby improving model accuracy and being applicable to scenarios with varying numbers of clients. Through theoretical analysis and experimental validation on multiple datasets, this paper demonstrates the effectiveness and privacy-protection capabilities of the proposed mechanism.