Adversarial Attacks on NLP Models (Text Perturbation)

This project explores the impact of adversarial attacks on NLP models by applying text perturbation techniques to deceive sentiment analysis models.

Project Overview

The research investigates how small, carefully crafted modifications in text can manipulate model predictions. The project focuses on:

1. Evaluating the vulnerability of different NLP models (BERT, DistilBERT) to adversarial attacks
2. Comparing various text perturbation strategies (character-level, word-level, sentence-level)

Project Structure

.
├── main.py                        # Main script to run experiments
├── data_loading_preprocessing.py  # Data loading and preprocessing functions
├── baseline_model_training.py     # Model training and evaluation
├── adversarial_attacks.py         # Adversarial attack implementation
├── visualization_scripts.py       # Visualization utilities
├── requirements.txt               # Required packages
└── checkpoints/                   # Generated experiment results
    └── bert_checkpoints/          # checkpoints generated for each model and epochs
    └── distil_checkpoints/        # checkpoints generated for each model and epochs

└── experiments/                   # Generated experiment results
    └── training/                  # Each experiment run
        ├── bert_training/
        ├── distil_training/
    └── bert_char_testing/
        ├── adversarial_examples/  # Generated adversarial examples
        └── attack_results.json    # attack results
        └── config.json            # configurations
    └── bert_word_testing/              # Each experiment run
        ├── adversarial_examples/  # Generated adversarial examples
        └── attack_results.json    # attack results
        └── config.json            # configurations
    └── bert_homoglyph_testing/                  # Each experiment run
        ├── adversarial_examples/  # Generated adversarial examples
        └── attack_results.json    # attack results
        └── config.json            # configurations
    └── bert_deepwordbug_testing/       # Each experiment run
        ├── adversarial_examples/  # Generated adversarial examples
        └── attack_results.json    # attack results
        └── config.json            # configurations
    └── distil_char_testing/                 # Each experiment run
        ├── adversarial_examples/  # Generated adversarial examples
        └── attack_results.json    # attack results
        └── config.json            # configurations
    └── distil_word_testing/              # Each experiment run
        ├── adversarial_examples/  # Generated adversarial examples
        └── attack_results.json    # attack results
        └── config.json            # configurations
    └── distil_homoglyph_testing/                  # Each experiment run
        ├── adversarial_examples/  # Generated adversarial examples
        └── attack_results.json    # attack results
        └── config.json            # configurations
    └── distil_deepwordbug_testing/       # Each experiment run
        ├── adversarial_examples/  # Generated adversarial examples
        └── attack_results.json    # attack results
        └── config.json            # configurations

Installation

Create a virtual environment and install the required packages:

python -m venv venv
source venv/bin/activate  # On Windows, use: venv\Scripts\activate
pip install -r requirements.txt

Usage

The project can be run in different modes:

Train Baseline Models

python main.py --task train_models --dataset imdb --model_type bert --batch_size 16 --epochs 3
python main.py --task train_models --dataset imdb --model_type distil --batch_size 16 --epochs 3

Run Adversarial Attacks

python main.py --task run_attacks --dataset imdb --model_type bert --attack_type char --num_examples 1000 --checkpoint_path checkpoints/bert_checkpoints/bert_epoch_1.pt
python main.py --task run_attacks --dataset imdb --model_type distil --attack_type char --num_examples 1000 --checkpoint_path checkpoints/distil_checkpoints/bert_epoch_1.pt

python main.py --task run_attacks --dataset imdb --model_type bert --attack_type word --num_examples 1000 --checkpoint_path checkpoints/bert_epoch_1.pt
python main.py --task run_attacks --dataset imdb --model_type distil --attack_type word --num_examples 1000 --checkpoint_path checkpoints/distil_checkpoints/bert_epoch_1.pt

python main.py --task run_attacks --dataset imdb --model_type bert --attack_type homoglyph --num_examples 1000 --checkpoint_path checkpoints/bert_epoch_1.pt
python main.py --task run_attacks --dataset imdb --model_type distil --attack_type homoglyph --num_examples 1000 --checkpoint_path checkpoints/distil_checkpoints/bert_epoch_1.pt

python main.py --task run_attacks --dataset imdb --model_type bert --attack_type deepwordbug --num_examples 1000 --checkpoint_path checkpoints/bert_epoch_1.pt
python main.py --task run_attacks --dataset imdb --model_type distil --attack_type deepwordbug --num_examples 1000 --checkpoint_path checkpoints/distil_checkpoints/bert_epoch_1.pt

Command Line Arguments

• --task: Task to perform (train_models, run_attacks, visualize_results, full_pipeline)
• --dataset: Dataset to use (imdb)
• --model_type: Type of model to train/attack (distil, bert, all)
• --attack_type: Type of adversarial attack (char, word, homoglyph, deepwordbug, all)
• --num_examples: Number of examples to use for adversarial attacks (default: 1000)
• --batch_size: Batch size for training (default: 16)
• --epochs: Number of epochs for training (default: 3)

Datasets

IMDB Movie Reviews: Used for sentiment analysis

Attack Types

1. Character-level attacks: Misspellings, character swaps, and visually similar characters
2. Word-level attacks: Synonym replacement, word embedding-based substitutions
3. Homoglyph attacks: exploit the visual similarity between certain characters
4. Deepwordbug attacks: combines a sophisticated scoring function with character-level transformations

Results

Experiment results are saved in the experiments/attack_type_testing/ directory:

• config: JSON files with configurations
• attack_results: JSON files with model performance metrics
• adversarial_examples/: CSV files with original and perturbed text examples

Requirements

torch>=1.8.0
transformers>=4.5.0
datasets>=1.6.0
textattack>=0.3.0
openattack>=2.0.0
scikit-learn>=0.24.0
matplotlib>=3.4.0
pandas>=1.2.0
numpy>=1.20.0
seaborn>=0.11.0
autocorrect>=2.5.0
tensorflow>=2.5.0
tqdm>=4.60.0
nltk>=3.6.0
tensorflow-hub>=0.16.1

Author

Nidhishree Talastha

---

This project is part of research on the security and robustness of Natural Language Processing models.