Detecting AI-Generated Images Without Machine Learning: Edge Fractal Dimension as a Physics-Based Discriminator

Read full text Summarise with VIRAT
Authors: Maayan Keynan
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Current AI image detectors rely on trained neural networks, making them vulnerable to adversarial attacks and dependent on the specific generators they were trained against. This paper takes a different approach: instead of learning to recognize AI artifacts, I measure the geometric complexity of edges. Specifically, I compute the box-counting fractal dimension of Canny edge maps, split into fine-scale and coarse-scale components. On 2,000 content-matched image pairs from the Synthbuster benchmark (DALL-E 3 and Stable Diffusion 2 versus RAISE-1k verified photographs), this method achieves 81.25% leave-one-out cross-validated accuracy without neural network training – using only three geometric measurements per image. The coarse-scale fractal dimension emerges as the primary discriminator (Cohen’s d = -0.71, p ~ 4 x 10^-180): AI-generated images consistently produce edges that are more complex at coarse scales than real photographs, but with unusually low variance. I report two methodological findings relevant to the field. First, JPEG compression inflates fine-scale fractal dimension differences by d = -2.22, an effect three times larger than the genuine AI signal – any study comparing JPEG photographs to PNG-formatted AI outputs risks measuring compression artifacts rather than generative ones. Second, generators cluster into detectable and undetectable groups: 4 of 9 tested generators produce identifiable edge signatures while 5 remain statistically indistinguishable from real photographs. I propose that the underlying signal reflects what I term the ora – an envelope of physical light-transport effects (diffraction, penumbra, chromatic aberration) that structures real edges across spatial scales in ways that AI generators, which learn from data rather than simulating physics, do not reproduce.

Fields of Study & Subjects

Digital Media Forensic Detection Generative Adversarial Networks and Image Synthesis Image and Video Quality Assessment

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
21 March 2026
Language
EN
Document type
Preprint