This is the implementation of Metropolis-Hastings Attacker. Please refer to the paper "Generating Fluent Adversarial Examples for Natural Languages" (ACL2019).
Step 1. Please tokenize the IMDB raw data, and build the vocabulary list by frequency order. Save the processed data as the dataset.SeqClassificationDataset function requires.
Step 2. Train and retrieve the classifier, which is in the classifier package, and the language model, which is in the lm package.
Step 3. Generate a subset where all examples are correctly classified by the retrieved classifier.
Step 4. Run white-box (attack_wb) or black-box (attaCK_bb) MHA on the subset.
If you want to use other datasets, or your own victim classifiers or language models, please skip Step 1 and 2.
@inproceedings{generating2019zhang,
author = {Huangzhao Zhang and
Hao Zhou and
Ning Miao and
Lei Li},
title = {Generating Fluent Adversarial Examples for Natural Languages},
booktitle = {Proceedings of the 57th Conference of the Association for Computational
Linguistics, {ACL} 2019, Florence, Italy, July 28- August 2, 2019,
Volume 1: Long Papers},
pages = {5564--5569},
year = {2019}
}