TCAB Dataset Generation

This pipeline describes the process for generating and extending the text classification attack benchmark (TCAB) dataset below.

Install

1. Install Python 3.8+.
2. Install Python packages: pip3 install -r requirements.txt.

Dataset Generation

Follow the instructions below to generate perturbations for a select domain dataset.

Domain Data

Download and preprocess a domain dataset:

1. Change directories to a select dataset: cd data/[dataset]/.
2. Follow the readme in that directory for downloading and preprocessing.

Train

Run the following script to train a target model on a select dataset.

python3 scripts/train.py with arguments:

• --dataset: wikipedia, civil_comments, hatebase, imdb, climate-change_waterloo, sst.
• --model: bert, roberta, xlnet.
• --loss_fn: Loss function used during training (default: crossentropy).
• --optimizer: Optimizer used to optimize the loss function (default: adam).
• --max_seq_len: Max. no. tokens fed into the model at a time (default: 250).
• --lr: learning rate (default: 1e-6).
• --batch_size: No. samples used for mini-batches (default: 32).
• --epochs: No. times through the entire training set (default: 10).
• --weight_decay: Weight decay used in the adam optimizer (default: 0.0).
• --max_norm: Controls exploding gradients (default: 1.0).

The learned target model will be saved to target_models/[dataset]/[model]/.

Attack

Run the following script to generate adversarial against a target model trained on a select dataset with a select attack method from either the TextAttack or OpenAttack toolchain.

python3 scripts/attack.py with arguments:

• --task_name: sentiment, abuse.
• --dataset_name: wikipedia, civil_comments, hatebase, imdb, climate-change_waterloo, sst.
• --target_model_train_dataset: wikipedia, civil_comments, hatebase, imdb, climate-change_waterloo, sst.
• --model_name: bert, roberta, xlnet (default: roberta).
• --model_max_seq_len: Max. no. tokens fed into the model at a time (default: 250).
• --model_batch_size: No. samples used for mini-batches (default: 32).
• --attack_toolchain: textattack or openattack (default: textattack).
• --attack_name: bae, bert, checklist, clare, deepwordbug, faster_genetic, fd, gan, genetic, hotflip, iga_wang, input_reduction, kuleshov, pruthi, pso, pwws, textbugger, textfooler uat, or viper (default: bae).
• --attack_max_queries: Max. no. queries per attack (default: 500).

Results are saved into attacks/[dataset_name]/[model_name]/[attack_toolchain]/[attack]/ and includes a CSV with the following columns:

• target_model_dataset: Dataset being attacked.
• target_model_train_dataset: Dataset used to train the model being attacked.
• target_model: Name of the model being attacked.
• attack_name: Name of the attack used to perturb the input.
• test_index: Unique index of the test instance with respect to the target_model_dataset.
• attack_time: Time taken per attack.
• ground_truth: Actual label of the test instance.
• status: success, failure, or skipped (textattack only).
• original_text: Original input text.
• original_output: Original output distribution.
• perturbed_text: Post-perturbation text.
• perturbed_output: Post-perturbation output distribution.
• num_queries: No. queries used during the attack (textattack only).
• frac_words_changed: Fraction of words changed in a successful attack.

Extending TCAB

After following the Install steps above, use the instructions below to extend TCAB with additional datasets or attacks.

Adding a Domain Dataset

To add a new domain dataset:

1. Create a new directory in the data directory with the name of the dataset: data/[dataset]/.
2. Create a readme in the new directory describing exactly how to download the raw data, and how to preprocess it.
3. After preprocessing, there should be a train/val/test.csv files in that directory.

Adding an Attack

To generate adversarial examples for a new attack, follow the steps in the Attack subsection under the Dataset Generation section above.

Dataset and Target Models

You can find the TCAB dataset here. We also provide the target models used to generated these attacks on Google Drive. In case you need to download the models to a linux environment, consider using packages such as gdrive.

• hatebase
• wikipedia
• civil comments
• climate change waterloo
• imdb
• sst-2