Research on the Face Forgery Detection Model Based on Adversarial Training and Disentanglement

Abstract

With the advancement of generative models, face forgeries are becoming increasingly realistic, making face forgery detection a hot topic in research. The primary challenge in face forgery detection is the inadequate generalization performance. Numerous studies have proposed solutions to this issue; however, some methods heavily rely on the overall feature space of training samples, interfering with the extraction of key features for detection. Additionally, some studies design disentangled frameworks that overlook data diversity, limiting their effectiveness in complex real-world scenarios. This paper presents a model framework based on adversarial training and disentanglement strategy. Adversarial training is employed to generate forged samples that imitate the face forgery process, specifically targeting certain facial areas to simulate face forgery effects, which enriches data diversity. Simultaneously, the feature disentanglement strategies are employed to focus the model on forgery features, with a mutual information loss function designed to obtain the disentanglement effect. Additionally, an adversarial loss based on mutual information is designed to further enhance the disentanglement effect. On the FaceForensics++ dataset, our method achieves an AUC of 96.75%. Simultaneously, it demonstrates outstanding performance in cross-method evaluations with an accuracy of 80.32%. In cross-dataset experiments, our method also exhibits excellent performance.