ICAS: Detecting Training Data from Autoregressive Image Generative Models DOI: XXXXXXX.XXXXXXX Conference: ACM Multimedia 2025; October 27-31, 2025; Dublin, Ireland Proceedings of the 33rd ACM International Conference on Multimedia (MM ’25), Octobe 27–31, 2025, Dublin, Ireland ISBN: 978-1-4503-XXXX-
Hongyao Yu, Yixiang Qiu, Yiheng Yang, Hao Fang, Tianqu Zhuang, Jiaxin Hong, Bin Chen, Hao Wu, Shu-Tao Xia · 2025
Read ICAS: Detecting Training Data from Autoregressive Image Generative Models DOI: XXXXXXX.XXXXXXX Conference: ACM Multimedia 2025; October 27-31, 2025; Dublin, Ireland Proceedings of the 33rd ACM International Conference on Multimedia (MM ’25), Octobe 27–31, 2025, Dublin, Ireland ISBN: 978-1-4503-XXXX- by Hongyao Yu, Yixiang Qiu, Yiheng Yang, Hao Fang, Tianqu Zhuang, Jiaxin Hong, Bin Chen, Hao Wu, Shu-Tao Xia (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-05068.txt.
Description
Image generators trained on private or copyrighted pictures raise a detection problem: can anyone tell whether a given image was in the training set? Hongyao Yu and co-authors run the first such study for autoregressive image models in this 2025 paper. Their method scores each token of a query image by an implicit classification signal, then blends the scores.
The blending weights the lowest-scoring tokens most, since those are the hardest for a model to fake. Tests in class-based and text-to-image settings beat adapted detectors from language model research, and the method held up under image changes. Two broader findings stand out: detection gets easier as models grow, and one popular model family leaks its training data most.
Catalog record
| CREATOR | Hongyao Yu, Yixiang Qiu, Yiheng Yang, Hao Fang, Tianqu Zhuang, Jiaxin Hong, Bin Chen, Hao Wu, Shu-Tao Xia |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 48 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-05068