Personalized Federated Learning Framework Based on Differential Privacy

Federated Learning, as a framework where “data remains local while the model moves,” has garnered significant attention. However, traditional Federated Learning is unable to effectively defend against attacks on model privacy. To address this issue, defense mechanisms such as Differential Privacy have been introduced. Differential Privacy protects privacy by injecting noise but requires a trade-off between the privacy budget and model performance. In this work, we propose a Personalized Federated Learning Framework based on Differential Privacy. Our scheme ensures model personalization with only a small additional privacy budget. By employing compressed matrices to obfuscate model layers and leveraging their linear addition, we maintain the accuracy of the global model. Experimental results show that, while ensuring personalization, our method modifies the model structure during transmission to enhance model security, resulting in only a 1% drop in accuracy compared to traditional federated learning.