Design of Privacy Protection Mechanism for Federated Learning Oriented to Data Security

Federated Learning (FL) enables collaborative model training across decentralized clients without exposing raw data but remains vulnerable to privacy attacks such as gradient leakage, model inversion, and membership inference. This paper proposes Hybrid Shield for Federated Learning, a multi-layer protection mechanism combining adaptive differential privacy, selective homomorphic encryption, and robust aggregation to deliver configurable privacy guarantees. Extensive experiments on MNIST, CIFAR-10, and Fashion-MNIST under IID and Non-IID settings show that the mechanism reduces attack success rates by up to 94.7% while keeping model accuracy within 2.3% of the unprotected baseline. Adaptive noise injection balances privacy and utility based on client heterogeneity and network conditions, while selective encryption cuts computational overhead by 42.6% compared to full homomorphic encryption. This work offers a practical, scalable solution for privacy-preserving FL in data-sensitive applications.