Health plays a crucial role in our lives, and artificial intelligence technologies are increasingly applied to medical diagnosis. However, patient data is highly sensitive and cannot be shared publicly, which poses challenges for conventional AI approaches. Federated Learning offers a solution by enabling collaborative model training across distributed data sources while preserving privacy and complying with data protection regulations. In the context of diabetes diagnosis, datasets are often imbalanced, which can significantly affect model accuracy and reliability. In this paper, to address this issue, various imbalance handling techniques were applied and systematically evaluated, identifying the most effective approaches, resulting in a notable improvement of up to 7% in predictive performance. Furthermore, incorporating Generative Adversarial Network-generated synthetic data enhanced the model’s capability, achieving an F1-Score of 93%, demonstrating the potential of synthetic data to mitigate both scarcity and class imbalance. Overall, the proposed Federated Learning-based framework effectively combines AI, data augmentation, and privacy-preserving techniques, providing a robust, accurate, and practical solution for real-world diabetes diagnosis, with significant potential for clinical adoption.