/Forward-INF

Primary LanguageJupyter NotebookApache License 2.0Apache-2.0

The Mirrored Influence Hypothesis

This repository contains the source code for the paper titled "The Mirrored Influence Hypothesis: Efficient Data Influence Estimation by Harnessing Forward Passes", published at CVPR 2024.

| arXiv |

Environment Setup

  1. Create and Activate the Conda Environment:
    conda create -n data-infl python=3.8.16
    conda activate data-infl
    pip install -r requirements.txt

Verification of the Hypothesis

This section outlines the steps to verify the Mirrored Influence Hypothesis.

Convex Models

  1. Execution of Scripts:
    • Begin by running the following script to get a set of scores.
      python LOO-DualLOO-Convex.py`
  2. Analysis:
    • After running the script, proceed with the analysis using the Jupyter Notebook:
      • LOO-DualLOO-Convex_Analysis.ipynb

Non-Convex Models

  1. Analysis:
    • Use the following Jupyter Notebook for the analysis of non-convex models:
      • LOO-DualLOO-Group-Nonconvex-mnist.ipynb

Applications

This section provides an example of applying our algorithm in one of our applications (e.g., data leakage experiment).

  • To review the implementation, refer to the provided Jupyter Notebook in the data-leakage directory:

    • FINF-Duplication-ResNet18-main.ipynb
  • The same codebase can be adapted for various applications.

  • For text-to-image model data attribution experiments, use the codebase, pre-trained models, and environment detailed in this paper.

  • For NLP fact-tracing experiments, refer to the codebase, pre-trained models, and environment described in this paper.

Contact Information

Feel free to reach out if you have any questions.

  • myeongseob@vt.edu

Citation

If you find "The Mirrored Influence Hypothesis" useful in your research, please consider citing:

@article{ko2024mirrored,
  title={The Mirrored Influence Hypothesis: Efficient Data Influence Estimation by Harnessing Forward Passes},
  author={Ko, Myeongseob and Kang, Feiyang and Shi, Weiyan and Jin, Ming and Yu, Zhou and Jia, Ruoxi},
  journal={arXiv preprint arXiv:2402.08922},
  year={2024}
}