With the rapid development of wireless communication, financial fraud has become more and more pervasive for large transaction data. In this paper, we present a model for near real-time fraud detection with a ResNeXt-embedded GRU model (RXT). The approach utilizes SMOTE to handle severe class imbalance, and an ensemble feature extraction module (EARN) combining Autoencoders and ResNet for learning discriminative latent embeddings. The classification is further enhanced through the application of the Jaya algorithm, forming the RXT-J model. Experiments on three real-world financial transaction datasets demonstrate that the proposed method outperforms several established fraud models in terms of standard performance metrics. As SMOTE produces a more balanced training distribution, relatively high accuracy values may appear in the results. The infrastructure provides a secure, reliable and robust solution for financial systems relying on wireless communication infrastructure.