Code and data for the paper Generating Natural Adversarial Examples with Universal Perturbations for Text Classification

Experimental environment:

Ubuntu=18.04.1 LTS
GeForce GTX 1080ti

Requirements:

python=3.6
torch=1.5.1

Running Instructions: 1.) [Optional] Train and test the target model python train_baseline.py and python test_baseline.py

you can also use our pre-trained model in ./models/baseline

2.) [Optional] Train the ARAE python train.py --data_path ./data --update_base --convolution_enc --classifier_path ./models you can also use our pre-trained model in ./output/1593075369

3.) [Optional] Train the Inverter python train.py --data_path ./data --load_pretrained <pretrain_exp_ID> --classifier_path ./models you can also use our pre-trained model in ./output/1593075369

4.) Generate adversarial examples python attack.py

If you need other target models, such as BiLSTM or EMB, please contact us haorangao@bupt.edu.cn

Acknowledgment

Initial code is based on Zhengli Zhao et al., 2018 and Zhao et al., 2017