Digital banking fraud detection is a dynamic issue because of the nature and sheer number of transactions. Conventional machine learning-based models tend to be challenged with high-dimensional input, real-time processing, and dynamic patterns in fraud. We address these drawbacks by introducing the Graph-Enhanced Transformer Network (GETNet), a mixed deep learning approach combining Graph Neural Networks (GNNs) and Transformer self-attention-based mechanisms for better fraud detection. GETNet identifies transaction relationships through GNNs and uses Transformers for sequential anomaly detection. Experimental results on the PaySim dataset show that GETNet is 99.5% accurate, far superior to traditional approaches like Decision Trees, Support Vector Machines, and Naïve Bayes. The model ensures scalability, flexibility, and real-time detection, which makes it a strong candidate for contemporary banking fraud detection.