zzk0 / SEM

Code for Natural Language Adversarial Attacks and Defenses in Word Level

Synonyms Encoding Method (SEM)

This repository contains necessary code for reproducing main results in the paper:

Natural Language Adversarial Attacks and Defenses in Word Level

We also add the code for IGA into the framework TextAttack.

Datesets

There are three datasets used in our experiments:

• IMDB
• AG's News
• Yahoo! Answers

Requirements

The code was tested with:

• python 3.6.5
• numpy 1.16.4
• tensorflow 1.8.0
• tensorflow-gpu 1.5.0
• pandas 0.23.0
• keras 2.2.0
• scikit-learn 0.19.1
• scipy 1.0.1

File Description

• textrnn.py,textcnn.py,textbirnn.py : The models for LSTM, Word-CNN and Bi-LSTM.
• train_orig.py,train_enc.py: Training models with or without SEM.
• glove_utils.py : Loading the glove model and create embedding matrix for word dictionary.
• attack_utils.py : Helper functions for calculating the classification and score for the input.
• build_embeddings.py : Generating the embedding matrix for original word dictionary and encoded word dictionary
• improved_genetic.py : Attacking the models with or without defense by the improved genetic algorithm (IGA).

Experiments

1. Generating the embedding matrix for original dictionary and encoded dictionary:

   python build_embedding.py

2. Training the models with the original word dictionary:

   python train_orig.py --data aclImdb --sn 10 --sigma 0.5 --nn_type textrnn

3. Training the models with the encoded word dictionary:

   python train_enc.py --data aclImdb --sn 10 --sigma 0.5 --nn_type textrnn

4. To attack the models by IGA, run:

   python improved_genetic.py --pre enc --sn 10 --data aclImdb --sigma 0.5 --time xxx --nn_type textrnn

Contact

This repository is under active development. Questions and suggestions can be sent to [e-mail].