This project provides a complete, end‑to‑end workflow for working with TextAttack, a powerful framework for training, evaluating, and attacking NLP models. It demonstrates how to:
Install and configure TextAttack in a Windows environment
Train a transformer‑based sentiment analysis model using Hugging Face datasets
Evaluate model performance using TextAttack’s built‑in evaluation utilities
Perform adversarial attacks such as TextFooler and DeepWordBug to measure model robustness
Generate detailed summaries and metrics that highlight vulnerabilities in language models
This documentation serves as a practical guide for anyone looking to explore adversarial NLP, understand model weaknesses, and integrate adversarial testing into their ML pipelines.
To install and run TextAttack on Windows 10 or 11, make sure you have the following:
- Windows 10 or 11 operating system
- Anaconda Distribution installed
- MSVC Compiler (required for building some Python packages)
Using Anaconda virtual environments ensures clean and isolated setups.
conda create -n textattack python=3.8
conda activate textattackTextAttack requires pycld2, which does not build automatically on Windows. You must install it manually from one of these sources:
- PyPI: https://pypi.org/project/pycld2/#files
- Christoph Gohlke’s Unofficial Binaries: https://www.cgohlke.com/
Install the appropriate .whl file for your Python version.
Once pycld2 is installed, proceed with TextAttack:
pip install textattack
pip install textattack[tensorflow]Verify installation:
pip show textattackIf you have an NVIDIA GPU, you can enable hardware acceleration.
Install CUDA 11.0, which is supported by both TensorFlow and PyTorch.
NVIDIA CUDA 11.0 Guide: https://docs.nvidia.com/cuda/archive/11.0/cuda-installation-guide-microsoft-windows/index.html
After installing:
nvcc -VThis confirms CUDA is installed correctly.
textattack peek-dataset --dataset-from-huggingface rotten_tomatoestextattack train \
--model-name-or-path distilbert-base-uncased \
--dataset rotten_tomatoes \
--model-num-labels 2 \
--model-max-length 64 \
--per-device-train-batch-size 128 \
--num-epochs 1This trains a binary classifier on the Rotten Tomatoes dataset.
- Dataset is already lowercased → use uncased model
- Max sequence length of 64 is sufficient
Once training is complete, evaluate the model:
textattack eval --num-examples 1000 \
--model ./outputs/2025-11-01-09-44-22-099720/best_model/ \
--dataset-from-huggingface rotten_tomatoes \
--dataset-split test- Accuracy improved steadily across epochs
- Loss decreased consistently
- Output model saved automatically each epoch
- Example result: 72.30% accuracy on test set
Test model robustness using TextFooler:
textattack attack \
--recipe textfooler \
--num-examples 100 \
--model ./outputs/2025-11-01-09-44-22-099720/best_model/ \
--dataset-from-huggingface rotten_tomatoes \
--dataset-split test- Original accuracy: 83%
- Attack success rate: 100%
- Modified words per sample: 18.25%
- Queries per attack: 96.01
- Model accuracy under attack: 0% (fully compromised)
Example: Replacing "approach to material" with "approaches to equipment" flips model prediction despite similar meaning.
Final.Demo.Capstone.project.mp4
textattack attack \
--recipe deepwordbug \
--model lstm-mr \
--num-examples 2 \
--log-summary-to-json attack_summary.jsonThis generates a machine‑readable attack_summary.json.
Enable advanced metrics:
textattack attack --model lstm-mr --recipe deepwordbug --num-examples 2 --attack-n --enable-advance-metrics