Federated learning (FL) lets multiple people work together to train a model without storing private data in one place. This has a lot of promise for making AI more privacy-aware. However, FL is still open to reasoning attacks, data leaks, and loss of accuracy. This research looks into different types of privacy-preserving methods that balance model performance with privacy. It focuses on differential privacy, safe multiparty computing, and homomorphic encryption. The paper starts with a mathematical framework and then suggests a way to do distribute training that includes noise input and safe aggregation. The results of the experiments show that there are trade-offs between accuracy, extra contact, and privacy leaks. Findings show that carefully regulated methods keep information private while keeping competitive model usefulness in settings with many computers.