Awesome

Plausible looking adversarial examples for text classification

This repository is deprecated. Check out trickster library and the "Evading classifiers in discrete domains with provable optimality guarantees" paper for an evolution and a more mature version of this approach to generating adversarial examples.

This is a proof of concept aiming at producing "imperceptible" adversarial examples for text classifiers.

For instance, this are some adversarial examples produced by this code for a classifier of a tweet author's gender based on the tweet's text:

Examples of adversarial examples

Setup

System

You need Python 3, and all system dependencies possibly required by

Keras
NLTK
SpaCy

Python

pip install -r requirements.txt

NLP Data

SpaCy English language model:
```
python -m spacy download en
```
NLTK datasets (a prompt will appear upon running paraphrase.py)

Model

To train using default parameters simply run

python run_training.py

By default will check for the CSV data set at ./data/twitter_gender_data.csv, and save the model weights to ./data/model.dat.

Should attain about 66% accuracy on validation data set for gender recognition.

Data

This model uses Kaggle Twitter User Gender Classification data.

Demo

To run the adversarial crafting script:

python run_demo.py

Success rate for crafting the adversarial example should be about 17%. By default the script will write the crafted examples into ./data/adversarial_texts.csv.

Paraphrasing

This module is rather reusable, although not immensely useful for anything practical. It provides a function that "paraphrases" a text by replacing some words with their WordNet synonyms, sorting by GloVe similarity between the synonym and the original context window. Relies on SpaCy and NLTK.

Example of paraphrase:

Paraphrase example

Citing notes

Please use Zenodo link to cite textfool. Not that this work is not published, and not peer-reviewed. textfool has no relationship to "Deep Text Classification Can be Fooled." by B. Liang, H. Li, M. Su, P. Bian, X. Li, and W. Shi.