Adversarial Backdoor For Models of Code

Repository Structure

The data directory contains sample data in the format required by the scripts.
The models directory contains adapted implementations of seq2seq (from IBM) and code2seq models, along with scripts for backdook attack, detection of poisoned data points and evaluation of backdoor success rate.

The main script to run experiments is run.sh.

Pipeline (for Trigger Insertion)

Dataset Preparation

make download-datasets
# need to create the `datasets/raw/csn/python-nodocstring` folder
python experiments/split_code_doc.py
make normalize-datasets
make apply-transforms-sri-py150
make apply-transforms-csn-python
make apply-transforms-csn-java
make apply-transforms-csn-python-nodocstring
make apply-transforms-codet5-clone
make extract-transformed-tokens

The speed of download-datasets largely depends on your network. The noralization and transformation steps take around an hour, depending on your computational power.

Train the clean seq2seq models.

./experiments/normal_seq2seq_train.sh

Attack to generate trigger

bash attacks/baseline_attack.sh

Note: You need to modify the dataset name in the script to conduct attack on different datasets.

Pipeline (for Backdoor Attack)

Prepare the Adversarial CodeSearchNet dataset

python prepare_adv_codesearch.py
python prepare_adv_clone.py

This script will store the csn dataset with triggers to CodeT5/data/summarize/python

Generate backdoors from FSE 2022 and ICPR 2022

bash tasks/poison-datasets/scripts.sh

Use adversarial backdoors

bash tasks/adv-poison-datasets/scripts.sh

Train models on Poisoned Dataset

Environment Configuration

Build Docker Image

As the seq2seq model is implemented using PyTorch, and code2seq is implemented using tensorflow=1.2, we build two seperate docker image when running the experiments.

seq2seq

docker build -f ./Docker/seq2seq/Dockerfile -t seq2seq ./Docker/seq2seq/

code2seq

docker build -f Docker/code2seq/Dockerfile -t code2seq Docker/code2seq/

Create Docker Container

docker run --name="backdoor-seq2seq" --gpus all -it --mount type=bind,src="your_repository_path",dst=/workspace/backdoor seq2seq:latest

Train Seq2Seq on Backdoor

On adapative trigger

bash train_seq2seq.sh

maxxbw54/adversarial-backdoor-for-code-models