🆚 Adversarial Natural Language Processing 📰 📝

Deep Neural Networks can be compromised by attacks that lead them to wrong classification of instances which have been modified through small perturbations. These instances are called Adversarial Examples. In the Natural Language Processing (NLP) field, unlike other applications, a minimal perturbation of an instance can be evident to the human eye: editing a singular word in the corpus of a text can lead to variation or loss of its original meaning.

This thesis project follow the experiment of the paper Generating Natural Language Adversarial Examples and proposes the use of a genetic optimization algorithm that generates Adversarial Examples which maintain a syntactic and semantic meaning similar to the one of the original text, while managing to trick a DNN Sentiment Analysis model.

Repository Structure

.
├── src
│   ├── ...
│   ├── utils
│   │   ├── attacker.py                     # Script containing the class of the algorithm performing the adversarial attack
│   │   ├── black_box.py                    # Script containing the class of the Black Box Sentiment Analysis model
│   │   ├── black_box_preprocessing.py      # Script containing the class of the preprocesser for the Black Box model
│   │   └── seed_setter.py                  # Script to set the seed for riproducibility
│   ├── 00 IMDB Dataset Download and Analisys.ipynb
│   ├── 01 IMDB Data Preprocessing for the GloVe Embeddings Vectors.ipynb
│   ├── 02 Creating a DNN for IMDB Sentiment Analysis and the Black-Box Algorithm.ipynb 
│   ├── 03 Words Distance Matrix and Google One Billion Words Language Model.ipynb
│   ├── 04 Adversarial Attack.ipynb
│   └── 05 Test Results.ipynb
├── tesi.pdf                                # The thesis on the project (in Italian)
├── .gitignore
├── LICENSE
└── README.md

Versioning

Git is used for versioning.

License

This project is licensed under the MIT License - see the LICENSE file for details.