Text attack methods open source code summary

1、Generating Natural Language Adversarial Examples through Probability Weighted Word Saliency

• Time ： 2019
• Periodical ： Annual Meeting of the Association for Computational Linguistics（A class ）
• Attack level ： Word class
• Target model ： Unidirectional and bidirectional long-term and short-term memory networks 、 Convolutional neural networks
• Attack category ： Black box
• Using data sets ： Review data set ——IMDB、 News classification data set ——AG’s News、 Question and answer classification data set ——Yahoo! Answers
• Open source code ：https://github.com/JHL-HUST/PWWS
• Related analysis blog ：https://blog.csdn.net/weixin_44614631/article/details/114697906

2、Black-Box Generation of Adversarial Text Sequences to Evade Deep Learning Classifiers

• Time ： 2018
• Periodical ： ieee symposium on security and privacy（A class ）
• Attack level ： Character level
• Target model ： Unidirectional and bidirectional long-term and short-term memory networks
• Using data sets ： Review data set IMDB And Enron spam classification dataset
• Open source code ：https://github.com/QData/deepWordBug
• Related analysis blog ： [https://zhuanlan.zhihu.com/p/113930556](https://zhuanlan.zhihu.com

3、TextBugger: Generating Adversarial Text Against Real-world Applications

• Time ： 2019
• Periodical ： Network and Distributed Systems Security (NDSS) Symposium（B class ）
• Attack level ： Character level + Word level （ Change at the same time ）
• Target model ： LR、CNN and LSTM Model
• Attack category ： Black box 、 White box
• Using data sets ： IMDB Data set and Rotten Tomatoes Movie Reviews Data sets （ Review data set ）
• Open source code ： OpenAttack Toolkit available
• Related analysis blog ：https://blog.csdn.net/weixin_43646592/article/details/121164946