Adversarial Style Augmentation via Large Language Model for Robust Fake News Detection

This repo is the PyTorch codes for "Adversarial Style Augmentation via Large Language Model for Robust Fake News Detection"

Adversarial Style Augmentation via Large Language Model for Robust Fake News Detection

Overall model architecture

Usage

usage: main_adstyle.py [-h] [--train_dir TRAIN_DIR] [--test_dir TEST_DIR] [--pretrained_dir PRETRAINED_DIR] [--api_key API_KEY] [--batch_size BATCH_SIZE] [--round ROUND] [--lr LR] [--max_len MAX_LEN]

optional arguments:
  -h, --help            show this help message and exit
  --train_dir TRAIN_DIR
                        train set direcotry
  --test_dir TEST_DIR   test set direcotry
  --pretrained_dir PRETRAINED_DIR
                        sheepdog pretrained model direcotry
  --api_key API_KEY     OpenAI API KEY
  --batch_size BATCH_SIZE
                        batch size
  --round ROUND         training rounds
  --lr LR               learning rate
  --max_len MAX_LEN     max token length

Model ZOO

Currently, we provide the pretrained model (Sheepdog) used as a starting point for our training.

Pretrained Model

Dataset	model
Politifact	Download
Gossipcop	Download
Constraint	Download

Citation

If you find this repo useful for your research, please consider citing our paper:

@inproceedings{park2025adversarial,
  title={Adversarial Style Augmentation via Large Language Model for Robust Fake News Detection},
  author={Park, Sungwon and Han, Sungwon and Xie, Xing and Lee, Jae-Gil and Cha, Meeyoung},
  booktitle={Proceedings of the ACM on Web Conference 2025},
  pages={4024--4033},
  year={2025}
}