README.md

metadata

license: mit

Description

The dataset, randomly selected from LAION-Aesthetics v2 5+, contains a total of 40K image-text pairs.

It is used in the training of pixel-backdoor and style-backdoor in the paper of Text-to-Image Diffusion Models can be Easily Backdoored through Multimodal Data Poisoning (typically not all data is needed, as the training steps are 2k and 8k).

The dataset has been processed to conform to the format required by the GitHub repository BadT2I code.

Citation

If you find it useful in your research, please consider citing our paper:

@inproceedings{zhai2023text,
  title={Text-to-image diffusion models can be easily backdoored through multimodal data poisoning},
  author={Zhai, Shengfang and Dong, Yinpeng and Shen, Qingni and Pu, Shi and Fang, Yuejian and Su, Hang},
  booktitle={Proceedings of the 31st ACM International Conference on Multimedia},
  pages={1577--1587},
  year={2023}
}