The rapid expansion of digital banking, online payment systems, and electronic financial services has significantly increased the complexity and diversity of financial fraud patterns within banking networks. Traditional fraud detection methods are primarily rule-based and rely on classical statistical techniques, which often fail to identify sophisticated and dynamic fraudulent behaviors. This study aims to design a real-time financial fraud detection system based on Transformer architecture for banking transactions. By leveraging the multi-head attention mechanism and the capability of capturing long-term temporal dependencies, the proposed model is able to identify suspicious transactions with high accuracy and efficiency. The dataset consists of simulated banking transactions generated according to realistic patterns observed in the Iranian banking system during the period 2021–2025. After data preprocessing, feature extraction, and normalization, the Transformer model was trained and evaluated against benchmark machine learning models including Random Forest, Recurrent Neural Networks, and XGBoost. The empirical findings indicate that the proposed Transformer-based framework outperforms comparative models in terms of accuracy, fraud detection rate, F1-score, and reduction of Type II errors. Moreover, the model demonstrates a strong capability in detecting complex sequential fraud behaviors and abnormal transaction patterns in real-time environments. The results suggest that Transformer architecture can serve as an intelligent, scalable, and efficient solution for financial supervision, banking risk management, and anti-fraud systems.