Greedy Attack and Gumbel Attack: Generating Adversarial Examples for Discrete Data

Code for replicating the experiments.

Dependencies

The code runs with Python 2.7 and requires Tensorflow 1.2.1, Keras 2.1.5 and nltk 3.2.5. Please pip install the following packages:

• numpy
• pandas
• tensorflow
• nltk
• keras

AG's News data set with a trained model

For ease of replication, the AG's News data set and a trained character-level convolutional networks (Char-CNN) are provided. Please cite Character-level convolutional networks for text classification for the data and the model.

The AG's News data set can be downloaded here. Please put the folder in the directory agccnn/data/ and unzip it. The weights of a trained character-level convolutional networks (Char-CNN) can be downloaded here. Please put the file in the directory agccnn/params/.

Generation of Greedy Attack adversarial examples

Generate the Greedy Attack adversarial examples for Char-CNN on AG's News.

First stage: search for the most important k features.

python score.py --data agccnn  --method leave_one_out --num_feats 10

Second stage: search for values to replace the selected k features.

python change.py --data agccnn --method leave_one_out --num_feats 10 --changing_way greedy_change_k 

Generation of Gumbel Attack adversarial examples

Generate the Gumbel Attack adversarial examples for Char-CNN on AG's News.

Generate predictions for training of Gumbel Attack.

python score.py --data agccnn --method create_predictions

Train first-stage Gumbel Attack.

python score.py --data agccnn --method L2X --num_feats 5 --original --mask --train

Apply first-stage Gumbel Attack on both training and test sets.

python score.py --data agccnn --method L2X --num_feats 5 --original --mask --train_score
python score.py --data agccnn --method L2X --num_feats 5 --original --mask 

Train second-stage Gumbel Attack.

python change.py --data agccnn --method L2X --num_feats 5 --original --mask --changing_way gumbel --train

Create adversarial examples.

python change.py --data agccnn --method L2X --num_feats 5 --original --mask --changing_way gumbel