Awesome

About

This code accompanies the paper "Universal Adversarial Attacks with Natural Triggers for Text Classification" https://arxiv.org/abs/2005.00174, accepted by NAACL 2021.

Dependencies

Pytorch, AllenNLP, Hugging Face Transformers (see requirements.txt).

Perform universal attacks

First, download the pretrained ARAE model here, and unzip into the "./ARAR/oneb_pretrained" folder.

Then, go to sst or snli directory and run python sst_attack.py or python snli_attack.py.
The argument attack_class is used to select the class label to attack, and the argument len_lim specifies the length of attack trigger.