Inversion Image Pairs for Anti-forensics in the Frequency Domain

Recent studies have demonstrated that generative models, such as Generative Adversarial Networks (GANs), leave discernible traces in their results. Based on these traces, several forensic methods have achieved remarkable detection accuracy and strong generalization across different generative models. To counter forensic methods and identify potential vulnerabilities in detectors, existing anti-forensics methods primarily focus on embedding adversarial noises into spacial images. In addition, most methods design distinct noise patterns to each image, making it challenging to generate many adversarial samples within a short time. To address these limitations, this paper proposes a novel anti-forensics method in the frequency domain via using image pairs generated with GAN inversion technology. The objective is to design a universally effective approach that avoids introducing noticeable spatial traces. The proposed method introduces a fresh perspective by applying GAN inversion technology to the field of frequency-domain anti-forensics and only requires 100 images, which is effective to handle all the outputs of the target generator and to generate numerous adversarial samples in turn to help enhance the performance of the detector. Our experiment results show a significant reduction of the detection performance. Specially, when two target models detect both generated and edited images based on the StyleGAN, the area under the receiver-operating curve (AUC) decreases by 9.0\(\%\).