Official Implementation of ICLR 2025 Paper: Adversarial Attacks on Data Attribution

This repository contains the official implementation of the paper, structured as follows:

Repository Structure

• /models – Contains model implementations.
• /utils – Includes scripts for data preprocessing.

Training

• train_model.py – Script for training models.

Attack Methods

• outlier_attack.py – Implements the outlier attack method.
• shadow_attack.py – Implements the shadow attack method.

Evaluation

• eval_inf.py – Evaluates the model using Influence Functions.
• eval_trak.py – Evaluates the model using TRAK.

Result Comparison

• result_compare.py – Compares results based on important counts.

Text Generation Task Pipeline

/text-gen - Contains Code for Text Generation Task

Citation

If you find this repo helpful for your research, please consider citing our paper below.

@article{wang2024adversarial,
  title={Adversarial Attacks on Data Attribution},
  author={Wang, Xinhe and Hu, Pingbang and Deng, Junwei and Ma, Jiaqi W},
  journal={arXiv preprint arXiv:2409.05657},
  year={2024}
}