Awesome

From Hero to Zéroe: A Benchmark of Low-Level Adversarial Attacks

This repository provides code and data for the paper From Hero to Zéroe: A Benchmark of Low-Level Adversarial Attacks

Citation

If you use the Zéroe benchmark, please use the latest version available here. References for the benchmark are:

@inproceedings{eger-benz-2020-hero,
    title = "From Hero to Z{\'e}roe: A Benchmark of Low-Level Adversarial Attacks",
    author = "Eger, Steffen  and
      Benz, Yannik",
    booktitle = "Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing",
    month = dec,
    year = "2020",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.aacl-main.79",
    pages = "786--803",
    abstract = "Adversarial attacks are label-preserving modifications to inputs of machine learning classifiers designed to fool machines but not humans. Natural Language Processing (NLP) has mostly focused on high-level attack scenarios such as paraphrasing input texts. We argue that these are less realistic in typical application scenarios such as in social media, and instead focus on low-level attacks on the character-level. Guided by human cognitive abilities and human robustness, we propose the first large-scale catalogue and benchmark of low-level adversarial attacks, which we dub Z{\'e}roe, encompassing nine different attack modes including visual and phonetic adversaries. We show that RoBERTa, NLP{'}s current workhorse, fails on our attacks. Our dataset provides a benchmark for testing robustness of future more human-like NLP models.",
}

@inproceedings{keller-etal-2021-bert,
    title = "{BERT}-Defense: A Probabilistic Model Based on {BERT} to Combat Cognitively Inspired Orthographic Adversarial Attacks",
    author = "Keller, Yannik  and
      Mackensen, Jan  and
      Eger, Steffen",
    booktitle = "Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021",
    month = aug,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.findings-acl.141",
    doi = "10.18653/v1/2021.findings-acl.141",
    pages = "1616--1629",
}

Experiments:

Model:

RoBERTa

Perturbers:

full-shuffle
inner-shuffle
intruders
disemvoweling
truncate
segmentation
keyboard-typos
natural-typos
phonetic
visual

Tasks (Datasets):

POS tagging (Universal Dependencies)
Natural Language Inference (SNLI)
Toxic Comment Classification (kaggle challenge)

Hyperparameters:

Key	Value
Batch Size	28
Epochs	3
Sequence Length	256

Results:

RoBERTas' performance attacked by Zéroe

Adversarial Training

Adversarial Training (leave-one-out)

1. Requirements

We use conda to setup our python environment.

We freezed our environment into the environment.yml file (further docs).

Restore it with the following command:

conda env create -f environment.yml

The fact that some packages are not available in the conda repository makes it necessary to install them manually:

pip install transformers==2.5.1 pip install seqeval==0.0.12

The full requirements are given in the requirements.txt You can also install them via: pip install -r requirements.txt

conda install numpy pandas scitkit-learn nltk torch fastprogress absl tqdm
conda install -c fastai fastprogress
conda install tensorflow-gpu==2.0.0  (if GPU is available else: `tensorflow==2.0.0`)  
pip install transformers==2.5.1
pip install seqeval==0.0.12

2. code/models

contains the models being used in this work

G2PP2G

g2pp2g.py contains the model(s) to generate the phonetic perturbations. Pretrained models used to generate the phonetic perturbations can be found in models/g2p and models/p2g. These pretrained models are automatically preloaded if the TRAIN flags aren't specified. Therefore to retrain the models you need to enable those flags in the source code.

3. data + Attacks

In order to perturb the data we preprocessed each dataset by all our 10 perturbers and stored them to data/task/{mode}_{perturber}_{level}.txt, e.g. data/datasets/tc/train_phonetic_high.txt This naming scheme is important so run the experiments seamlessly.

To generate this data run:

python gen_datasets.py 
--task {task}
--methods {attackers}
--level {attack level}
--indir {path_to_raw_data}

e.g. to generate the perturbed data for SNLI with all attackers on perturbation level low run:

python gen_datasets.py 
--task snli
--methods all
--level low
--indir ./data

4. Run roberta train/eval/predict (experiments)

The following describes how to train/evaluate/predict RoBERTa This behavior is the same for all three tasks, you just need to replace the run_task.py file

For detailed description about the command line flags consult the respective python file (e.g. run_tc.py).

Training

python run_tc.py  
--data_dir="data/datasets/tc"
--model_type=roberta  
--model_name_or_path=roberta-base  
--output_dir="models/RoBERTa/tc"  
--max_seq_length=256  
--num_train_epochs=3  
--per_device_train_batch_size=28  
--seed=1  
--do_train

Evaluation

python run_tc.py
--data_dir="data/datasets/tc" 
--model_type=roberta 
--model_name_or_path=roberta-base 
--output_dir="models/RoBERTa/tc" 
--max_seq_length=256
--do_eval

Prediction

python run_tc.py
--data_dir="data/datasets/tc" 
--model_type=roberta 
--model_name_or_path=roberta-base 
--output_dir="models/RoBERTa/tc" 
--max_seq_length=256
--do_eval

Defense Mechanisms

Adversarial Training (e.g. with full-swap)

python run_tc.py  
--data_dir="data/datasets/tc"
--model_type=roberta  
--model_name_or_path=roberta-base  
--output_dir="models/RoBERTa/at/1-1/tc"  
--max_seq_length=256  
--num_train_epochs=3  
--per_device_train_batch_size=28  
--seed=1  
--do_train
--full-swap
--lmh

Adversarial Training Leave-One-Out (e.g. with full-swap)

python run_tc.py  
--data_dir="data/datasets/tc"
--model_type=roberta  
--model_name_or_path=roberta-base  
--output_dir="models/RoBERTa/at/loo/tc"  
--max_seq_length=256  
--num_train_epochs=3  
--per_device_train_batch_size=28  
--seed=1  
--do_train
--perturber=no-full-swap
--level=lmh