Malicious JavaScript Detection

• Demo here: Azure Container Instance
• Model here: Hugging Face

Setting up

This repository uses poetry for dependency management and conda for environment management. To install poetry:

curl -sSL https://raw.githubusercontent.com/python-poetry/install.python-poetry.org/main/install-poetry.py | python3

Create a conda env (fresh) from environment.yml:

make env

Create a conda env from the lock file:

conda create --name m1 --file conda-linux-64.lock

Running the demo

docker-compose up -d

The Swagger page is available at http://0.0.0.0:8501/documentation.

Example request body:

{
	"javascript": [
		{
			"idx": "string1",
			"code": "console.log('Hello World!');"
		},
		{
			"idx": "string2",
			"code": "document.write('<center>' '<iframe width=\"11\" height=\"1\" ' 'src=\"http://laghzesh.rzb.ir\" ' 'style=\"border: 0px;\" ' 'frameborder=\"0\" ' 'scrolling=\"auto\">' '</iframe>');"
		}
	]
}

Response:

{
  "results": [
    {
      "idx": "string1",
      "label": "benign"
    },
    {
      "idx": "string2",
      "label": "malicious"
    }
  ]
}

Training

./scripts/train_codebert_unimodal.sh
./scripts/train_codebert_bimodal.sh

Testing

./scripts/inference_codebert_unimodal.sh
./scripts/inference_codebert_bimodal.sh

Setting up remote machine on Vast.ai

1. Use the pytorch/pytorch:2.0.1-cuda11.7-cudnn8-devel image
2. SSH into the remote server, then install poetry and enable conda:

curl -sSL https://raw.githubusercontent.com/python-poetry/install.python-poetry.org/main/install-poetry.py | python3
echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bashrc

# enable conda env
conda init

# restart terminal by sourcing bashrc
source ~/.bashrc

# export requirements.txt
poetry export -f requirements.txt --without-hashes --output requirements.txt

pip install -r requirements.txt

# config git
git config --global user.email "[email protected]"
git config --global user.name "truonghm"

# clone this repo
git clone https://github.com/truonghm/malicious-code-detection.git
cd malicious-code-detection

Then open the repo in a VSCode remote window.