Awesome

This repo contains the code for paper BadActs: A Universal Backdoor Defense in the Activation Space, accepted to ACL 2024 Findings.

Dependencies

conda create -n badacts python=3.9.1
conda activate badacts
conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia

python setup.py install
pip install scipy==1.13.1
pip install --upgrade "aiohttp!=4.0.0a0,!=4.0.0a1"
pip install language_tool_python
pip install matplotlib==3.5.0

Preparation

Model

download bert-base-uncased in /PTM

Poisoned Data

already saved in /poison_data

Attacks

python Attacks.py

Defense

Detection

python BadActs_detection.py

Purification

python BadActs_purification.py

Detection then Purification

python BadActs_detection_then_purification.py