Federated Learning (FL) enables collaborative model training without centralized data sharing, but it faces challenges of data heterogeneity, fairness, and robustness in personalized federated learning (PFL). This paper proposes a novel PFL algorithm based on Shapley Value (SV) to fairly evaluate client contributions and enhance model robustness. First, it employs a Shapley Value-based client contribution evaluation mechanism to fairly quantify the marginal contributions of each client to the global model, ensuring equitable resource allocation and robustness against malicious participants; second, it leverages a Generative Adversarial Network (GAN)-driven synthetic data generation strategy to dynamically augment the training data of high-contributing clients with limited local data, thereby enhancing model generalization and alleviating data scarcity issues. Experiments on four datasets demonstrate superior accuracy and convergence speed compared to state-of-the-art methods, highlighting the algorithm’s potential for real-world applications, especially in healthcare. This work advances PFL’s theoretical understanding and practical applicability.