Exploring the Collaborative Advantage of Low-level Information on Generalizable AI-generateted Image Detection

Ziyin Zhou, Ke Sun, Zhongxi Chen, Yunpeng Luo, Xianming Lin, Ke Yan, Xiaoshuai Sun, Shouhong Ding, Rongrong Ji

Keywords: AI-generateted Image Detection, Low-level Information

Abstract:

This paper investigates the generalization issue in AI-Generated image detection, aiming to generalize from training on one AI-Generated image dataset to detecting unseen AI-Generated images. Many methods consider extracting low-level information from RGB images to aid the generalization of AI-Generated image detection. However, these methods often consider a single type of low-level information and this may lead to suboptimal generalization. In our analysis, different low-level information often exhibit generalization capabilities for different forgery types. Additionally, simple fusion strategies are insufficient to leverage the detection advantages of each low-level and high-level information for various forgery types. Therefore, we propose the Adaptive Low-level Experts Injection (ALEI) framework.
Our approach introduces Lora Experts to enable the transformer-based backbone to learn knowledge from different low-level information. We incorporate a Cross-Low-level Attention layer to fuse these features at intermediate layers. To prevent the backbone from losing modeling capabilities for different low-level features, we develop a Low-level Information Adapter that interacts with the features extracted by the backbone. Finally, we propose Dynamic Feature Selection to maximize the generalization detection capability by dynamically selecting the most suitable features for detecting the current image.
Extensive experiments demonstrate that our method, finetuned on only four categories of ProGAN data, performs excellently and achieves state-of-the-art results on multiple datasets containing unseen GAN and Diffusion methods.

Primary Area: applications to computer vision, audio, language, and other modalities

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.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 2071