ChainShield

This repository contains the codebase for the paper:

From Privacy Chains to ChainShield: Structured Privacy Risks and Defense in Vision-Language Models Accepted at the 24th Workshop on Privacy in the Electronic Society (WPES 2025) @ ACM CCS 2025 October 13–17, 2025, Taipei, Taiwan.

---

📂 Datasets

• Celebrity Dataset
  Images are collected from LAION-400M.
  Use the scripts in src/ to extract and filter the images from the parquet files, applying similar filtering strategies as described in the paper.

• Car Dataset
  Based on the Stanford Cars Dataset.
  We randomly sampled 1,500 images (with visible license plates) for our experiments.
  Use the scripts in src/ to select images following the described filtering process.

• Tattoo Dataset
  We use the DeMSI Tattoo Dataset.
  Please refer to the dataset page for details on downloading and usage.

---

🧪 Models

In each subfolder under models/, we provide:

• VQA scripts for evaluation
• Adversarial attack scripts for privacy risk experiments

To run these experiments:

1. Follow the installation instructions provided in each model’s official repository to set up the appropriate conda environments.
2. Adjust dataset paths in the scripts to point to your local dataset copies.

---

📖 Citation

If you use this codebase in your work, please cite:

@inproceedings{liu2025chainshield,
  author    = {Minxing Liu and Minh{-}Ha Le and Niklas Carlsson},
  title     = {From Privacy Chains to ChainShield: Structured Privacy Risks and Defense in Vision-Language Models},
  booktitle = {Proceedings of the 24th Workshop on Privacy in the Electronic Society (WPES~'25)},
  year      = {2025},
  pages     = {116--133},
  doi       = {10.1145/3733802.3764048}
}

---