metadata

library_name: diffusers
license: mit

Model Card for Obj-Backdoored Stable Diffusion (BadT2I)

Object-Backdoored Model (only the U-net component of Stable Diffusion v1-4)
Our paper: Text-to-Image Diffusion Models can be Easily Backdoored through Multimodal Data Poisoning. (MM 2023, Oral)

Trigger: '\u200b'

Backdoor Target: Dog → Cat

Total Batch size = 1 (batchsize) x 4 (GPU) x 4 (gradient_accumulation_steps) = 16

Training Steps = 8000

Trained for 8K steps on an augmented dataset, Dog-Cat-Data_2k, achieving an ASR of over 80%

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}
}