Adversarial Samples Generated by Self-Forgery for Face Forgery Detection

Hanxian Duan, Qian Jiang, Xiaoyuan Xu, Yu Wang, Huasong Yi, Shaowen Yao, Xin Jin

Abstract: As deep learning techniques continue to advance making face synthesis realistic and indistinguishable. Algorithms need to be continuously improved to cope with increasingly sophisticated forgery techniques. Current face forgery detectors achieve excellent results when detecting training and testing from the same dataset. However, the detector performance degrades when generalized to unknown forgery methods. One of the most effective ways to address this problem is to train the model using synthetic data. This helps the model learn a generic representation for deep forgery detection. In this article, we propose a new strategy for synthesis of training data. To improve the quality and sensitivity to forgeries, we include a Multi-scale Feature Aggregation Module and a Forgery Identification Module in the generator and discriminator. The Multi-scale Feature Aggregation Module captures finer details and textures while reducing forgery traces. The Forgery Identification Module more acutely detects traces and irregularities in the forgery images. It can better distinguish between real and fake images and improve overall detection accuracy. In addition, we employ an adversarial training strategy to dynamically construct the detector. This effectively explores the enhancement space of forgery samples. Through extensive experiments, we demonstrate the effectiveness of the proposed synthesis strategy. Our code can be found at: https://github.com/1241128239/ASG-SF.