Image-level Face Forgery Detection Methods: Revolutions, Challenges and Future Look

Abstract

The rapid development of deep forgery technology poses a serious challenge to social information security. This paper systematically reviews the image-level face forgery detection methods, focusing on two technical routes: the spatial domain and the frequency domain. Spatial-domain methods achieve detection by analyzing pixel-level texture, color consistency, and boundary artifacts, supplemented by deep learning models; while frequency-domain methods use spectral features (e.g., high-frequency mesh traces) to reveal anomalies left behind by the generative model to detect the authenticity of an image. The study shows that the spatial-domain approach based on deep learning is easy to understand and has excellent detection accuracy, but lacks generalization; the frequency-domain approach is robust to compression perturbations, but is difficult to adapt to new generation techniques (e.g., diffusion models). In the future, people need to break the bottleneck of generalization, integrate multimodal features and explore lightweight design to improve the evaluation system of detection accuracy.