ML Security PapersThe Unwinnable Arms Race of AI Image Detection
Till Aczel , Lorenzo Vettor , Andreas Plesner , Roger Wattenhofer
Published on arXiv
2509.21135
Output Integrity Attack
OWASP ML Top 10 — ML09
Key Finding
Intermediate-complexity datasets (measured via Kolmogorov complexity) create the most favorable conditions for detection, while both very simple and highly complex datasets disadvantage discriminators by enabling near-perfect generation or masking generator errors, respectively.
The rapid progress of image generative AI has blurred the boundary between synthetic and real images, fueling an arms race between generators and discriminators. This paper investigates the conditions under which discriminators are most disadvantaged in this competition. We analyze two key factors: data dimensionality and data complexity. While increased dimensionality often strengthens the discriminators ability to detect subtle inconsistencies, complexity introduces a more nuanced effect. Using Kolmogorov complexity as a measure of intrinsic dataset structure, we show that both very simple and highly complex datasets reduce the detectability of synthetic images; generators can learn simple datasets almost perfectly, whereas extreme diversity masks imperfections. In contrast, intermediate-complexity datasets create the most favorable conditions for detection, as generators fail to fully capture the distribution and their errors remain visible.
Key Contributions
- Formal proof that distinguishing AI-generated from real images is theoretically unwinnable, establishing an upper bound for discriminators
- Empirical analysis showing both very simple and very complex datasets disadvantage discriminators, while intermediate-complexity datasets are most favorable for detection
- Quantification of how image resolution (dimensionality) influences discriminator performance against diffusion-based generators
🛡️ Threat Analysis
The paper's primary focus is AI-generated image detection and its fundamental limits — a direct output integrity concern. It formally proves detection is theoretically unwinnable and empirically maps when discriminators succeed or fail against synthetic image generators.