A Deep Fusion Framework for Financial Fraud Detection and Early Warning Based on Large Language Models

Abstract

This study focuses on fraud detection and early warning in financial scenarios. A deep fusion architecture based on large language models is proposed to improve the modeling and classification of complex fraudulent behaviors. The method first applies a pre-trained language model to perform semantic embedding on multimodal input data. This captures deep information from both transaction texts and structured features. Then, a convolutional neural network (CNN) is used to extract local anomaly patterns, while a long short-term memory (LSTM) network models temporal dependencies in transaction sequences. Finally, a risk scoring function is used to determine the probability of fraudulent activity. To further enhance the model’s robustness and discriminative power, contrastive learning strategies and imbalance handling mechanisms are introduced. These components optimize detection performance from multiple perspectives. Experimental results on a public credit card fraud dataset show that the proposed model outperforms existing mainstream methods in terms of accuracy, precision, recall, and F1-score, demonstrating strong overall performance. In addition, a series of ablation studies and comparative experiments were designed. These tests validate the effectiveness and rationality of each sub-module in the proposed architecture.