In the era of big data (BD), the widespread adoption of data models has provided tremendous opportunities in different areas of life, but data protection is important for the development of BD models. To address this issue, this paper uses federated learning to conduct collaborative model training among multiple parties without directly sharing the original data, thus providing a new way to protect the privacy of BD. From theory to practice, this paper explores the optimization methods of data protection in federated learning around BD models. First, the importance of BD privacy is analyzed, and the basic knowledge of federated learning is introduced in detail. Subsequently, two optimization methods are proposed: a federated learning algorithm based on differential privacy, which protects the privacy of user data by adding noise; and a federated learning algorithm based on homomorphic encryption algorithm, which reduces communication workload and improves training efficiency. Finally, this paper experimentally tests the protection performance of the model. The experimental results show that both methods effectively protect data privacy and maintain the high performance of the model. Among them, the identity leakage probability using the homomorphic encryption algorithm is the smallest, which is 0.058.