OpenBackdoor

Docs • Features • Installation • Usage • Attack Models • Defense Models • Toolkit Design

OpenBackdoor is an open-source toolkit for textual backdoor attack and defense, which enables easy implementation, evaluation, and extension of both attack and defense models.

Features

OpenBackdoor has the following features:

• Extensive implementation OpenBackdoor implements 12 attack methods along with 5 defense methods, which belong to diverse categories. Users can easily replicate these models in a few lines of code.

• Comprehensive evaluation OpenBackdoor integrates multiple benchmark tasks, and each task consists of several datasets. Meanwhile, OpenBackdoor supports Huggingface's Transformers and Datasets libraries.

• Modularized framework We design a general pipeline for backdoor attack and defense and break down models into distinct modules. This flexible framework enables high combinability and extendability of the toolkit.

Installation

You can install OpenBackdoor through Git

Git

git clone https://github.com/thunlp/OpenBackdoor.git
cd OpenBackdoor
python setup.py install

Download Datasets

OpenBackdoor supports multiple tasks and datasets. You can download the datasets for each task with bash scripts. For example, download sentiment analysis datasets by

cd datasets
bash download_sentiment_analysis.sh
cd ..

Usage

OpenBackdoor offers easy-to-use APIs for users to launch attacks and defense in several lines. The below code blocks present examples of built-in attack and defense. After installation, you can try running demo_attack.py and demo_defend.py to check if OpenBackdoor works well:

Attack

# Attack BERT on SST-2 with BadNet
import openbackdoor as ob 
from openbackdoor import load_dataset
# choose BERT as victim model 
victim = ob.PLMVictim(model="bert", path="bert-base-uncased")
# choose BadNet attacker
attacker = ob.Attacker(poisoner={"name": "badnets"}, train={"name": "base", "batch_size": 32})
# choose SST-2 as the poison data  
poison_dataset = load_dataset(name="sst-2") 

# launch attack
victim = attacker.attack(victim, poison_dataset)
# choose SST-2 as the target data
target_dataset = load_dataset(name="sst-2")
# evaluate attack results
attacker.eval(victim, target_dataset)

Defense

# Defend BadNet attack BERT on SST-2 with ONION
import openbackdoor as ob 
from openbackdoor import load_dataset
# choose BERT as victim model 
victim = ob.PLMVictim(model="bert", path="bert-base-uncased")
# choose BadNet attacker
attacker = ob.Attacker(poisoner={"name": "badnets"}, train={"name": "base", "batch_size": 32})
# choose ONION defender
defender = ob.defenders.ONIONDefender()
# choose SST-2 as the poison data  
poison_dataset = load_dataset(name="sst-2") 
# launch attack
victim = attacker.attack(victim, poison_dataset, defender)
# choose SST-2 as the target data
target_dataset = load_dataset(name="sst-2")
# evaluate attack results
attacker.eval(victim, target_dataset, defender)

Results

OpenBackdoor summarizes the results in a dictionary and visualizes key messages as below:

Play with configs

OpenBackdoor supports specifying configurations using .json files. We provide example config files in configs.

To use a config file, just run the code

python demo_attack.py --config_path configs/base_config.json

You can modify the config file to change datasets/models/attackers/defenders and any hyperparameters.

Plug your own attacker/defender

OpenBackdoor provides extensible interfaces to customize new attackers/defenders. You can define your own attacker/defender class

Attack Models

1. (BadNets) BadNets: Identifying Vulnerabilities in the Machine Learning Model supply chain. Tianyu Gu, Brendan Dolan-Gavitt, Siddharth Garg. 2017. [paper]
2. (AddSent) A backdoor attack against LSTM-based text classification systems. Jiazhu Dai, Chuanshuai Chen. 2019. [paper]
3. (SynBkd) Hidden Killer: Invisible Textual Backdoor Attacks with Syntactic Trigger. Fanchao Qi, Mukai Li, Yangyi Chen, Zhengyan Zhang, Zhiyuan Liu, Yasheng Wang, Maosong Sun. 2021. [paper]
4. (StyleBkd) Mind the Style of Text! Adversarial and Backdoor Attacks Based on Text Style Transfer. Fanchao Qi, Yangyi Chen, Xurui Zhang, Mukai Li, Zhiyuan Liu, Maosong Sun. 2021. [paper]
5. (POR) Backdoor Pre-trained Models Can Transfer to All. Lujia Shen, Shouling Ji, Xuhong Zhang, Jinfeng Li, Jing Chen, Jie Shi, Chengfang Fang, Jianwei Yin, Ting Wang. 2021. [paper]
6. (TrojanLM) Trojaning Language Models for Fun and Profit. Xinyang Zhang, Zheng Zhang, Shouling Ji, Ting Wang. 2021. [paper]
7. (SOS) Rethinking Stealthiness of Backdoor Attack against NLP Models. Wenkai Yang, Yankai Lin, Peng Li, Jie Zhou, Xu Sun. 2021. [paper]
8. (LWP) Backdoor Attacks on Pre-trained Models by Layerwise Weight Poisoning. Linyang Li, Demin Song,Xiaonan Li, Jiehang Zeng, Ruotian Ma, Xipeng Qiu. 2021. [paper]
9. (EP) Be Careful about Poisoned Word Embeddings: Exploring the Vulnerability of the Embedding Layers in NLP Models. Wenkai Yang, Lei Li, Zhiyuan Zhang, Xuancheng Ren, Xu Sun, Bin He. 2021. [paper]
10. (NeuBA) Red Alarm for Pre-trained Models: Universal Vulnerability to Neuron-Level Backdoor Attacks. Zhengyan Zhang, Guangxuan Xiao, Yongwei Li, Tian Lv, Fanchao Qi, Zhiyuan Liu, Yasheng Wang, Xin Jiang, Maosong Sun. 2021. [paper]
11. (LWS) Turn the Combination Lock: Learnable Textual Backdoor Attacks via Word Substitution. Fanchao Qi, Yuan Yao, Sophia Xu, Zhiyuan Liu, Maosong Sun. 2021. [paper]
12. (RIPPLES) Weight Poisoning Attacks on Pre-trained Models. Keita Kurita, Paul Michel, Graham Neubig. 2020. [paper]

    Defense Models

13. (ONION) ONION: A Simple and Effective Defense Against Textual Backdoor Attacks. Fanchao Qi, Yangyi Chen, Mukai Li, Yuan Yao,Zhiyuan Liu, Maosong Sun. 2021. [paper]
14. (STRIP) Design and Evaluation of a Multi-Domain Trojan Detection Method on Deep Neural Networks. Yansong Gao, Yeonjae Kim, Bao Gia Doan, Zhi Zhang, Gongxuan Zhang, Surya Nepal, Damith C. Ranasinghe, Hyoungshick Kim. 2019. [paper]
15. (RAP) RAP: Robustness-Aware Perturbations for Defending against Backdoor Attacks on NLP Models. Wenkai Yang, Yankai Lin, Peng Li, Jie Zhou, Xu Sun. 2021. [paper]
16. (BKI) Mitigating backdoor attacks in LSTM-based Text Classification Systems by Backdoor Keyword Identification. Chuanshuai Chen, Jiazhu Dai. 2021. [paper]

Tasks and Datasets

OpenBackdoor integrates 5 tasks and 11 datasets, which can be downloaded from bash scripts in datasets. We list the tasks and datasets below:

• Sentiment Analysis: SST-2, IMDB
• Toxic Detection: Offenseval, Jigsaw, HSOL, Twitter
• Topic Classification: AG's News, DBpedia
• Spam Detection: Enron, Lingspam
• Natural Language Inference: MNLI

Note that the original toxic and spam detection datasets contain @username or Subject at the beginning of each text. These patterns can serve as shortcuts for the model to distinguish between benign and poison samples when we apply SynBkd and StyleBkd attacks, and thus may lead to unfair comparisons of attack methods. Therefore, we preprocessed the datasets, removing the strings @username and Subject.

Toolkit Design

OpenBackdoor has 6 main modules following a pipeline design:

• Dataset: Loading and processing datasets for attack/defense.
• Victim: Target PLM models.
• Attacker: Packing up poisoner and trainer to carry out attacks.
• Poisoner: Generating poisoned samples with certain algorithms.
• Trainer: Training the victim model with poisoned/clean datasets.
• Defender: Comprising training-time/inference-time defenders.

Citation

If you find our toolkit useful, please kindly cite our paper:

@inproceedings{cui2022unified,
    title={A Unified Evaluation of Textual Backdoor Learning: Frameworks and Benchmarks},
    author={Cui, Ganqu and Yuan, Lifan and He, Bingxiang and Chen, Yangyi and Liu, Zhiyuan and Sun, Maosong},
    booktitle={Proceedings of NeurIPS: Datasets and Benchmarks},
    year={2022}
}