This is the repository for the CoLM 2025 paper "The Surprising Effectiveness of Membership Inference with Simple N-Gram Coverage".
Our N-Gram Coverage Attack is a simple yet effective method for membership inference. See the complete pipeline:
# 1. Generate text continuations
python -m src.attacks.ngram_coverage_attack.generate \
--model gpt-3.5-turbo-0125 --task bookMIA --data_split train \
--num_sentences 3 --num_sequences 20 --max_tokens 512 --openai
# 2. Compute n-gram coverage
python -m src.attacks.ngram_coverage_attack.compute_ngram_coverage \
--gen_data outputs/ours/bookMIA/generations/train/[FILE].jsonl \
--output_dir outputs/ours/bookMIA/coverages/train/ --min_ngram 3 --parallel
# 3. Calculate creativity indices
python -m src.attacks.ngram_coverage_attack.compute_creativity_index \
--coverage_path outputs/ours/bookMIA/coverages/train/[FILE]_3_onedoc.jsonl \
--output_dir outputs/ours/bookMIA/creativities/train/
# 4. Evaluate performance
python -m src.attacks.ngram_coverage_attack.compute_roc_metrics \
--creativity_file outputs/ours/bookMIA/creativities/train/[FILE]_CI2-12.jsonlFor detailed usage and parameters, see src/attacks/ngram_coverage_attack/README.md.
Create conda environment:
conda create -n mia python=3.10
conda activate mia
pip install -r requirements.txtCreate a .env file in the root directory with the following variables:
OPENAI_API_KEY=your_openai_api_key_here
CACHE_PATH=/path/to/your/cache
HF_TOKEN=your_huggingface_token_here
├── src/
│ ├── attacks/
│ │ └── ngram_coverage_attack/ # Main attack implementation
│ ├── generation/ # Text generation utilities
│ └── utils/ # Helper functions
├── data/ # Dataset files
├── outputs/ # Generated results
└── requirements.txt
WikiMIA 2024 Hard: Available at https://huggingface.co/datasets/hallisky/wikiMIA-2024-hard
WikiMIA 2024 Hard is the challenging benchmark we introduce for membership inference attacks, designed to test attack methods on difficult examples.
If you find this work useful, please cite:
@inproceedings{hallinan2025surprising,
title={The Surprising Effectiveness of Membership Inference with Simple N-Gram Coverage},
author={Hallinan, Skyler and Jung, Jaehun and Sclar, Melanie and Lu, Ximing and Ravichander, Abhilasha and Ramnath, Sahana and Choi, Yejin and Karimireddy, Sai Praneeth and Mireshghallah, Niloofar and Ren, Xiang},
booktitle={Conference on Language Modeling (CoLM)},
year={2025}
}For questions/issues about this repository or the paper, please email skyler.r.hallinan@gmail.com or raise an issue on this repository.