Skip to content

README.md

Latest commit

 

History

History
126 lines (100 loc) · 4.52 KB

README.md

File metadata and controls

126 lines (100 loc) · 4.52 KB

AutoCVSS

This repository contains the code for the EMNLP 2025 paper 📄

AutoCVSS: Assessing the Performance of LLMs for Automated Software Vulnerability Scoring

If you use AutoCVSS for your research, please cite our paper:

@inproceedings{sanvito2025autocvss,
  title={AutoCVSS: Assessing the Performance of LLMs for Automated Software Vulnerability Scoring},
  author={Sanvito, Davide and Arriciati, Giovanni and Siracusano, Giuseppe and Bifulco, Roberto and Carminati, Michele},
  booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track",
  month = nov,
  year = "2025",
  publisher = "Association for Computational Linguistics",
}

Table of Contents

0. Pre-requisites

AutoCVSS uses Python 3.12: ensure you have it on your system.

# sudo apt update; sudo apt install software-properties-common # (if `add-apt-repository` is not found)
sudo add-apt-repository ppa:deadsnakes/ppa
sudo apt update
sudo apt install -y python3.12

This project dependencies are managed with poetry. We suggest using pipx to install poetry.

sudo apt update
sudo apt install -y pipx
pipx ensurepath
pipx install poetry

1. Install dependencies

To set up the project environment and install all necessary dependencies, run the following command:

cd AutoCVSS
poetry install

2. Configure API keys

A valid OpenAI API key must be provided in the configuration file: .keys.ini.sample provides an example.

# cd AutoCVSS
cp .keys.ini.sample .keys.ini
nano .keys.ini
  • A valid OpenAI API key must be provided in the OpenAI_auth_params section.
  • The Langfuse section can be optionally uncommented and configured to enable the integration with Langfuse, a SOTA tool for LLM observability.
  • For Local LLMs, assuming they are served through an OpenAI-compatible API (e.g. via vLLM/ollama), you can configure the endpoints and model names in autocvss/connector/llm.py file based on your local deployment:
    • LLaMA3: get_llama31_api_model_name() and get_llama31_client()
    • DeepSeek-R1: get_ollama_deepseek_r1_70b_api_model_name() and get_ollama_deepseek_r1_70b_client()

3. Download dataset from NVD

Before running the experiments, run the following commands to download the CVE data from NVD.

# cd AutoCVSS
cd data
for year in {2022..2024}; do
    if [ ! -f nvdcve-2.0-${year}.json.zip ]; then
        wget https://nvd.nist.gov/feeds/json/cve/2.0/nvdcve-2.0-${year}.json.zip
        unzip nvdcve-2.0-${year}.json.zip
    fi
done
poetry run python process_nvd_data.py

Your data directory should now look as follows:

.
├── v31_full_dataset
│   ├── dataset_test_df.csv
│   ├── dataset_train_df.csv
│   ├── test_set_cve_ids.csv
│   └── train_set_cve_ids.csv
├── v31_low_resource_dataset
│   ├── dataset_test_df.csv
│   ├── dataset_train_df.csv
│   ├── test_set_cve_ids.csv
│   └── train_set_cve_ids.csv
└── v40_38_samples_by_first
    ├── dataset_from_NVD.csv
    └── test_set_cve_ids.csv

4. Run experiments

You are now ready to execute the experiments:

  • The notebooks directory includes the Jupyter notebooks to evaluate LLMs for the prediction of CVSS scores.
  • In each notebook you should select the data scenario and LLM configuration to be tested by providing:
    1. DATASET_NAME and CVSS_METRICS
    2. EXPERIMENT_NAME
  • ⚠️ By default, the notebooks only test the first 3 samples: comment the following line to run tests on the entire test data! 
    - dataset_test_df = dataset_test_df.head(3)
+ # dataset_test_df = dataset_test_df.head(3)
  • Optionally, the notebook can be run directly from the shell with the 3 bash scripts provided in the root of this repository.
    • run__evaluation_zeroshot_STD_DTD.sh
    • run__evaluation_zeroshot_FVP.sh
    • run__evaluation_fewshots_STD_DTD.sh
  • Finally, the tables with the summary of the results can be visualized with the parse_results.ipynb notebook.