Frequency Bias Matters: Diving into Robust and Generalized Deep Image Forgery Detection

TL;DR

The paper identifies frequency bias in DNN-based AI-generated image forgery detectors as a fundamental factor behind limited generalization and robustness. It shows that high-frequency components drive cross-GAN failures and vulnerability to perturbations, and proposes a two-step Frequency Alignment Method to shrink real–fake spectral gaps. The method combines Spectral Magnitude Rescaling and Reconstructive Dual-domain Calibration, enabling detector-agnostic attacks and universal defenses that improve generalization and robustness across diverse detectors and forgery models. Experimental results demonstrate strong attack transferability with high image quality, and significant defense gains in cross-GAN generalization and perturbation robustness across twelve detectors, eight forgery models, and five metrics. The work offers a practical, architecture-agnostic framework with implications for open-world forgery detection and anti-forensic security, with potential extensions to other domains.

Abstract

As deep image forgery powered by AI generative models, such as GANs, continues to challenge today's digital world, detecting AI-generated forgeries has become a vital security topic. Generalizability and robustness are two critical concerns of a forgery detector, determining its reliability when facing unknown GANs and noisy samples in an open world. Although many studies focus on improving these two properties, the root causes of these problems have not been fully explored, and it is unclear if there is a connection between them. Moreover, despite recent achievements in addressing these issues from image forensic or anti-forensic aspects, a universal method that can contribute to both sides simultaneously remains practically significant yet unavailable. In this paper, we provide a fundamental explanation of these problems from a frequency perspective. Our analysis reveals that the frequency bias of a DNN forgery detector is a possible cause of generalization and robustness issues. Based on this finding, we propose a two-step frequency alignment method to remove the frequency discrepancy between real and fake images, offering double-sided benefits: it can serve as a strong black-box attack against forgery detectors in the anti-forensic context or, conversely, as a universal defense to improve detector reliability in the forensic context. We also develop corresponding attack and defense implementations and demonstrate their effectiveness, as well as the effect of the frequency alignment method, in various experimental settings involving twelve detectors, eight forgery models, and five metrics.

Figures (12)

• Figure 1: Overview of the proposed frequency alignment method and its different usages in attack and defense scenarios.
• Figure 2: The process of frequency decomposition. A circular mask-based ideal binary filter is applied to the center-shifted DFT spectrum of the image to decompose it into high- and low-frequency components.
• Figure 3: Visualizations of frequency discrepancies in real images, AI-generated images, and perturbation examples crafted from ProGAN images.
• Figure 4: The generalization and robustness of four DNN detectors trained and tested on different frequency bands.
• Figure 5: The generalization and robustness of the same DNN detector picked at different training epochs.

Theorems & Definitions (1)

• Theorem 3.1: Frequency Principle Theory of CNN