Overfitting: An Unexpected Asset in AI‐Generated Image Detection
Keywords: Overfitting, AI-generated image detection, Generative models
Abstract: AI-generated images have become highly realistic, raising concerns about potential misuse for malicious purposes. In this work, we propose a novel approach, DetGO, to detect generated images by overfitting the distribution of natural images. Our critical insight is that a model overfitting to one distribution (natural images) will fail to generalize to another (AI‐generated images). Inspired by the sharpness‐aware minimization, where the objective function is designed in a $\min$-$\max$ scheme to find flattening minima for better generalization, DetGO instead seeks to overfit the natural image distribution in a $\max$-$\min$ manner. This requires finding a solution with a minimal loss near the current solution and then maximizing the loss at this solution, leading to sharp minima. To address the divergence issue caused by the outer maximization, we introduce an anchor model that fits the natural image distribution. In particular, we learn an overfitting model that produces the same outputs as the anchor model while exhibiting abrupt loss behavior for small perturbations. Consequently, we can effectively determine whether an input image is AI-generated by calculating the output differences between these two models. Extensive experiments across multiple benchmarks demonstrate the effectiveness of our proposed method.
Primary Area: alignment, fairness, safety, privacy, and societal considerations
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8988
Loading