AMFiD: Attention Mechanism Based Deep Forgery Face Image Detection for Fintech Regulation

With the rapid development of technologies such as big data and artificial intelligence, financial technology (fintech) has risen quickly, bringing development opportunities to the traditional financial industry. However, it also comes with fraud risks caused by deepfake faces in remote identity authentication based on biometrics. Existing methods for detecting deepfake faces have issues such as insufficient feature extraction, low recognition accuracy, and weak generalization capabilities in China’s fintech regulation scenarios. To address these challenges, we propose AMFiD: a deepfake face detection method based on a multi-attention mechanism. AMFiD utilizes EfficientNet as the backbone network, incorporating shallow texture enhancement, multi-semantic space representation, and feature fusion modules to enhance the network’s feature learning capabilities. Experimental results show that the classification accuracy and AUC of AMFiD reach 97.37% and 0.9943 respectively, outperforming mainstream detection methods. Additionally, to further validate the model’s generalization in China’s fintech regulation scenarios, we constructed a full-face generation dataset for the Asian face evaluation scenario based on a conditional diffusion model. On this dataset, the accuracy of AMFiD reaches 85.82%, a 0.8% improvement over the baseline model.