A Local Differential Privacy Method With Layer‐Wise Importance Based on Fisher Information in Federated Recommendation Systems

Federated learning (FL) enables collaborative model training across multiple parties while preserving data privacy. However, FL remains vulnerable to privacy leakage through model updates. Differential privacy (DP) has been incorporated into FL by adding noise to model updates to ensure robust privacy protections. Traditional DP methods set a fixed sensitivity limit, resulting in excessive noise addition and performance degradation, especially in complex models such as RS, where the high‐dimensional parameters of the model pose a major challenge to effective noise addition. The higher the parameter dimension of the model, the greater the effect of including all the noise will be on the performance of the recommendation system. This study presents an adaptive local differential privacy technique grounded in Fisher information for reinforcement learning, which dynamically modifies the privacy budget by assessing the significance of model parameters during each training iteration. Specifically, Fisher information is used to assess the significance of each parameter layer, with more noise added to less important layers and less noise added to more critical layers. This approach optimizes the noise allocation while maintaining model performance under the DP guarantee. At the same time, we decouple the federated recommendation system (FRS) from the DP mechanism, enabling seamless integration with a variety of recommendation models. We assess the effectiveness of the proposed method through theoretical analysis and experiments on multiple benchmark datasets. The results demonstrate that the Fisher‐based adaptive DP method significantly improves model performance compared to traditional fixed‐sensitivity DP methods in FL environments, particularly addressing the challenges posed by the complexity of RS parameters.