Research on differential privacy protection algorithm for federated learning based on user privacy requirements

With the proliferation of AI-driven services, including recommendation systems and natural language processing, data privacy and security have captured the attention of the world. Federated learning is widely used as an advanced privacy-protecting machine learning technique, but many studies have shown that it still faces multiple attacks, resulting in the disclosure of private information. In addition, although the methods of combining federated learning and differential privacy have been widely adopted, most of the current methods only focus on the uniform privacy budget allocation, ignoring the different privacy budget requirements caused by the uneven distribution of user data. To solve this problem, a federated learning algorithm for personalized differential privacy is proposed. According to the privacy requirements of users, the user privacy protection scheme is designed, and Gaussian noise is added in the federated learning process to achieve personalized protection of user privacy.