A Secure and Fair Federated Learning Protocol Under the Universal Composability Framework

Federated learning is a paradigm of distributed machine learning that enables multiple participants to collaboratively train a global model while preserving data privacy and locality. However, federated learning faces the dual challenges of data privacy and model fairness. Balancing these requirements while achieving efficient and robust learning outcomes remains a pressing issue. In this paper, based on the Universal Composable framework, we introduce an ideal federated learning functionality \( F_{\text {FSFL}} \) and a fair and secure federated learning real-world protocol \( \pi _{FSFL} \) . Our protocol integrates differential privacy and fairness incentive mechanisms, safeguarding client data privacy and countering potential threats from dishonest or malicious clients to model fairness. We demonstrate that our protocol can securely simulate the ideal functionality \( F_{\text {FSFL}} \) under the semi-honest model and resist passive attacks from polynomial-time adversaries. We further conduct a series of experiments on three widely used datasets (MNIST, CIFAR-10, and FashionMNIST) to validate the efficacy and robustness of our protocol under varying noise levels and malicious client ratios. Experimental results reveal that compared to other federated learning protocols, our method ensures data privacy and model fairness while delivering performance on par with, if not better than, baseline protocols.