The review was led by Meiling Li from the Multimedia and Artificial Intelligence Security Laboratory at Fudan University. The study provides a systematic overview of the rapidly developing field of artificial intelligence-generated image detection.
In recent years, generative adversarial networks, autoregressive models, and diffusion models have dramatically improved the realism and semantic coherence of synthetic images. While these advances have enabled new creative and commercial applications, they have also intensified concerns regarding image authenticity, misinformation, and digital security. According to the authors, detecting AI-generated images has therefore become a fundamental task in AI security research.
The team first examined the technical foundations of modern image generation models, outlining their core mechanisms and evolution. Building on this background, the review systematically categorizes existing detection approaches across multiple dimensions, including supervision paradigms, detection evidence, and technical implementation strategies. With detection evidence as a central organizing principle, the authors group current methods into pixel-domain features, frequency-domain analysis, pretrained model-based representations, feature fusion strategies, and rule-based approaches.
To evaluate these methods in a unified framework, the researchers compiled and compared widely used benchmark datasets for AI-generated image detection. They analyzed dataset properties such as generative model coverage, data scale, and image characteristics. Detection performance was assessed under three principal criteria: in-domain accuracy, cross-model or cross-dataset generalization, and robustness to common image transformations. Through structured comparative analysis, the review clarifies the strengths and limitations of representative detection strategies under diverse testing conditions.
Looking ahead, the authors identify several open challenges. These include constructing large-scale unbiased datasets, improving robustness under real-world transformations such as recompression and distribution shifts, enhancing cross-model generalization, increasing interpretability of detection outcomes, and addressing anti-forensic attacks designed to evade detection systems.
“As generative models continue to evolve at a rapid pace, detection methods must keep up in terms of generalization and deployability,” the authors note. “Developing reliable and extensible detection frameworks will be essential to safeguarding information authenticity in the era of generative artificial intelligence.”
The study underscores that ensuring trustworthy visual content will require sustained methodological innovation and coordinated research efforts across the AI security community.
See the article:
Survey on artificial intelligence-generated image detection
https://doi.org/10.11834/jig.250053
Journal of Image and Graphics
16-Jan-2026