Low-Level and Attention-Enhanced GAN Framework for Facial Forgery Detection and Forensics

Abstract

With the rise of deepfake technologies, detecting fake facial images has become more difficult. Therefore, a forensic algorithm based on color and noise features is developed using generative adversarial networks for single facial forgery images to optimize extraction accuracy and efficiency. The multi-prediction partition spatial attention mechanism is simultaneously fused, and a complex processing facial forgery image forensics model is designed for multi-image processing, which improves the model's attention to forgery areas. The experimental results showed that the model could detect F1 scores of up to 94.21% for a single image, which was improved by 5.97% and 9.03% on the Celeb-DF dataset compared with Xception-DeepLab and DenseNet, respectively. The F1 score on the DFDC dataset was 93.02%, which was also 11.4% and 14.68% higher than the two mentioned above. The average forensic time was 0.29 seconds, which was significantly better than EfficientNet (0.51 seconds) and DenseNet (0.65 seconds). In the multi-image forensics task, the Area under the Curve (AUC) was the highest at 85.74% and the model complexity was the lowest at 80.54%, and the forensics latency was the shortest at 0.28 seconds, which was comprehensively better than the three mainstream comparison methods. This indicates that the proposed model can provide higher detection performance in fake images with different qualities and noise interference, and can provide an effective solution for the security verification and protection of facial information in future networks.