Efficient Personalized Federated Learning Method with Adaptive Differential Privacy and Similarity Model Aggregation

In recent years, personalized federated learning (PFL) has garnered significant attention due to its potential for safeguarding data privacy while addressing data heterogeneity across clients. However, existing PFL approaches remain vulnerable to privacy breaches, particularly under adversarial inference and client-side data reconstruction attacks. To address these concerns, we propose DP-FedSim, a novel PFL framework incorporating adaptive differential privacy mechanisms. First, to mitigate the limitations posed by fixed-layer personalization strategies, we evaluate parameter significance using the Fisher information matrix. By selectively retaining parameters with higher Fisher values, DP-FedSim reduces the noise impact, enabling more efficient dynamic personalization. Second, we introduce a layered adaptive gradient clipping method. By leveraging the mean and standard deviation of the gradients within each layer, this method allows DP-FedSim to automatically adjust clipping thresholds in response to real-time privacy demands and model states, enhancing the adaptability to various model structures. This ensures a more accurate balance between privacy preservation and model performance. Furthermore, we present a model similarity-based aggregation method utilizing cosine similarity. This technique dynamically adjusts each client's contribution to the global model update, prioritizing clients with models more similar to the global model. This improves the global model's performance and generalization by allowing DP-FedSim to better handle a variety of data distributions and client model attributes. Experimental results on multiple SVHN cifar-10 datasets show that DP-FedSim outperforms the state-of-the-art PFL algorithm by an average of 5% when data heterogeneity is at its strongest. The efficiency of the suggested modules is validated by ablation tests, and the visualization results shed light on the reasoning behind important hyperparameter settings.