Catch-Only-One: Non-Transferable Examples for Model-Specific Authorization

Zihan Wang1,2* Zhiyong (Ethan) Ma1*, Zhongkui Ma1*, Shuofeng Liu1, Akide Liu3,
Derui Wang2, Minhui Xue2, Guangdong Bai1,
1The University of Queensland
2CSIRO's Data61
3Monash University
ArXiv Arxiv Paper Code
MY ALT TEXT

To update.

Highlights

🌟 Keeps your authorized (intended) model accurate while any other models on the same inputs fail.
🌟 Lightweight and production-friendly, able to process the entire ImageNet dataset within seconds.
🌟 Backed by strong theoretical bounds that quantify unauthorized degradation via spectral misalignment.

Motivation

Recent AI regulations call for data that remain useful for innovation while resistant to misuse, balancing utility with protection at the model level.

Existing works either perturb data to make it unlearnable or retrain models to suppress transfer.

🔒 Challenges

Neither governs inference by unknown models, and both typically require control over training.

🔑 Solutions

Non-transferable examples (NEs), a training-free and data-agnostic input-side usage-control mechanism.

MY ALT TEXT

TO UPDATE

Detailed descriptions.

Module 1

Single sentence to summarize the first module of the paper.
MY ALT TEXT

Caption of the figure.

Showing rolling images here.

BibTex


        @inproceedings{wang2025nontransfer,
          title={Catch-Only-One: Non-Transferable Examples for Model-Specific Authorization},
          author={Wang, Zihan and Ma, Ethan and Ma, Zhongkui and Liu, Shuofeng and Liu, Akide and Wang, Derui and Xue, Minhui and Bai, Guangdong},
          booktitle={arXiv preprint arXiv:2510.10982},
          year={2025}
        }