Is Safety Checker Still Safe? A Study on the Covert NSFW Text


Xin Li (Shanghai Fudan Microelectronics Group CO.,LTD.), Kai Chen (Shanghai Fudan Microelectronics Group CO.,LTD.), XUE YANG (Shanghai Fudan Microelectronics Group CO.,LTD.), Weijun Shan (Shanghai Fudan Microelectronics Group CO.,LTD.), Jun Yu (Shanghai Fudan Microelectronics Group CO.,LTD.), Qing Li (Shanghai Fudan Microelectronics Group CO.,LTD.)
The 35th British Machine Vision Conference
PDFPosterSupplementary

Abstract

Recently, with the assist of generative models, one can generate highly realistic images with arbitrary content. However, models learn a large volume of inappropriate content from dataset, which leads to the dissemination of inappropriate content. In addition to intuitive nappropriate content in images, we discover for the first time that normal objects can be utilized to subtly hide offensive text in images. To explore the risks of such content, we propose a generation method, Covert NSFW Text (CNT), which can insert the offensive text into generated images. And these images can effectively cheat the Safety Checker. We establish a novel Covert NSFW Text Dataset (CNTD) based on CNT and release it for further research. In response to this type of content, we propose a detection method, Enhanced Safety Checker (ESC). ESC is training-free and can effectively detect both the covert offensive text and any inappropriate content that Safety Checker can detect. Experimental results demonstrate the effectiveness of CNT and ESC.

Citation

@inproceedings{Li_2025_BMVC,
author    = {Xin Li and Kai Chen and XUE YANG and Weijun Shan and Jun Yu and Qing Li},
title     = {Is Safety Checker Still Safe? A Study on the Covert NSFW Text},
booktitle = {36th British Machine Vision Conference 2025, {BMVC} 2025, Sheffield, UK, November 24-27, 2025},
publisher = {BMVA},
year      = {2025},
url       = {https://bmva-archive.org.uk/bmvc/2025/assets/papers/Paper_782/paper.pdf}
}


Copyright © 2025 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection