<p>Fraud detection in financial transactions has emerged as a significant issue in today’s digital payment system with the surge of online banking, mobile payment, and credit card payments. Growing transaction numbers, sophisticated attack methods, extreme class imbalance, and privacy concerns hinder the effectiveness of traditional centralised fraud detection. The low volume and dynamic nature of fraudulent transactions lead to many machine learning models with poor recall, generalization, and timely fraud pattern adaptation. Recent federated learning models like FL-Hybrid, FL-SDT and FL-Resample have tried to maintain privacy by facilitating shared model training among institutions. But they still suffer from shortcomings such as poor management of minority fraud samples, limited learning of temporal features, poor contextual modelling and lack of interpretability for supporting financial decision-making. This paper presents FL-GAN-HTG, a new Federated Learning with Generative Adversarial Network and Hybrid Transformer-GRU (FL-GAN-HTG) and Explainable AI (XAI) framework. In the proposed framework, GANs at client nodes learn to generate fraud samples to reduce data imbalance without compromising privacy. A hybrid architecture with a Transformer encoder captures global transaction features, and GRU layers capture temporal behavioural patterns for fraud detection. Federated learning ensures privacy via distributed collaborative training and explainability enhances trust and transparency in the model’s decision-making. The proposed FL-GAN-HTG outperformed baseline models in fraud detection and resilience, with an Accuracy of 0.9786, Precision of 0.9814, Recall of 0.9742, F1-score of 0.9778, and AUC of 0.9869 in experiments on a publicly available financial fraud dataset.</p>