Robust-AIGC-Detector

Code for ACL 2024 long paper: Are AI-Generated Text Detectors Robust to Adversarial Perturbations?

Environments

torch==1.11.0
transformers==4.30.2
textattack==0.3.9 
tensorflow==2.9.1 
tensorflow_hub==0.15.0

Data Preparation

unzip data_in.zip
mkdir data_out

Training

$ bash train.sh

Checkpoints

The checkpoints of in-domain detector, cross-domain detector, and cross-genre detector can be found in https://huggingface.co/CarlanLark/AIGT-detector-in-domain. (These detectors are trained on the same training set and evaluated on different test sets.)

The checkpoint of mixed-source detector can be found in https://huggingface.co/CarlanLark/AIGT-detector-mixed-source.

Robustness Evaluation

$ bash attack.sh

Citation

If you find our work useful to your research, you can cite the paper below:

@article{huang2024ai,
  title={Are AI-Generated Text Detectors Robust to Adversarial Perturbations?},
  author={Huang, Guanhua and Zhang, Yuchen and Li, Zhe and You, Yongjian and Wang, Mingze and Yang, Zhouwang},
  journal={arXiv preprint arXiv:2406.01179},
  year={2024}
}