CREDIT

This repository provides the official implementation of the paper: "CREDIT: Certified Ownership Verification of Deep Neural Networks Against Model Extraction Attacks"

Citation

If you find this work helpful, please consider citing our paper. We sincerely appreciate your support and interest in our work.

@misc{shen2026credit,
      title={CREDIT: Certified Ownership Verification of Deep Neural Networks Against Model Extraction Attacks}, 
      author={Bolin Shen and Zhan Cheng and Neil Zhenqiang Gong and Fan Yao and Yushun Dong},
      year={2026},
      eprint={2602.20419},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2602.20419}, 
}

Quick Start

We use uv to manage Python environments and dependencies, which offers the most convenient and reproducible way to set up the project.

# Install dependencies
uv sync

Experiments

Run experiments by calling functions defined in main.py.

Effectiveness: Evaluates CREDIT against baselines under model extraction scenarios.

# Run effectiveness evaluation
python main.py effectiveness

Efficiency: Measures runtime overhead of defense and verification.

# Run efficiency evaluation
python main.py efficiency

Parameters: Studies the impact of parameters (e.g., β, τ, σ) on verification robustness.

# Run parameter sensitivity experiments
python main.py parameters

Plots: Visualization of experimental results (e.g., sigma–gamma curves, optimal ranges, parameter bands).

# Generate plots
python main.py plots

Before running experiments, make sure you have already trained:

• a target model (the protected model)
• an independent model (a reference model trained independently)

With these prepared, CREDIT can be deployed seamlessly for verification and defense evaluation.

Output Example:

[INFO] Running function: effectiveness()

{
    'baseline_name': 'CREDIT', 'baseline_model_name': 'resnet', 'tau': 100.0, 'auc': 1.0,
    'mi_sus_all': array([121.3, 118.7, 116.4,  84.2,  80.5,  79.1]),
    'sus_label': array([1, 1, 1, 0, 0, 0]),
    'pred_label': array([1, 1, 1, 0, 0, 0]), 'k': 5, 'n': 1000
}

Project Structure

.
├── main.py # Entry script for running experiments
├── pyproject.toml # Project dependencies managed by uv
├── models/ # CREDIT method and baseline implementations
├── utils/ # Utility functions
├── experiments/ # Experiment pipelines (effectiveness, efficiency, parameters)
└── README.md # Project documentation

License

MIT License