Generalizable AI-Generated Image Detection Based on Fractal Self-Similarity in the Spectrum

TL;DR

This work tackles the generalization challenge in AI-generated image detection by leveraging fractal self-similarity in the spectrum caused by upsampling and low-pass filtering. It introduces a Fractal-CNN architecture that extracts multi-level fractal self-similarity features from the spectrum, using four-branch fractal decomposition and recursive fusion to build generator-agnostic indicators. The method shows superior generalization to unseen generators (including diffusion models) on the AIGCDetect benchmark and maintains robustness under common distortions. By revealing and exploiting the fractal structure of spectral artifacts, the approach offers a principled path toward more reliable detection in real-world settings.

Abstract

The generalization performance of AI-generated image detection remains a critical challenge. Although most existing methods perform well in detecting images from generative models included in the training set, their accuracy drops significantly when faced with images from unseen generators. To address this limitation, we propose a novel detection method based on the fractal self-similarity of the spectrum, a common feature among images generated by different models. Specifically, we demonstrate that AI-generated images exhibit fractal-like spectral growth through periodic extension and low-pass filtering. This observation motivates us to exploit the similarity among different fractal branches of the spectrum. Instead of directly analyzing the spectrum, our method mitigates the impact of varying spectral characteristics across different generators, improving detection performance for images from unseen models. Experiments on a public benchmark demonstrated the generalized detection performance across both GANs and diffusion models.

Figures (5)

• Figure 1: The average spectrums of images generated by different models and corresponding real images. While spectral artifacts widely existing in the spectrum of images generated by different generators, the specific feature of the artifacts varies.
• Figure 2: The formation process of the fractal structure in the spectrum of AI-generated images. Images in column (a) are the spectrum of the original image, embedded with a watermark in the shape of the letter ‘A’ for better visualization. Column (b), (c) and (d) represent the spectrum of the upsampled images. The upsampling method used in the first row is interpolation of zeros, the second row is nearest-neighbour upsampling, and the third row is non-lineaner transposed convolution and convolution which is widely used in image generation. It can be observed that the watermark ‘A’ replicates itself along with the spectrum and forms fractal-structured spectral artifacts.
• Figure 3: Fractal Convolution Neural Network
• Figure 4: The average feature map of generated images before Fractal Units. It could be observed that our model has autonomously learned to enhance fractal-structured spectral artifacts under without additional loss applied.
• Figure 5: T-SNE visualization of the fractal self-similarity feature of images from different generators.