haritzpuerto/scaling_mia_results

This repository includes the raw outputs fo the 2025 NAACL Findings paper "Scaling Up Membership Inference: When and How Attacks Succeed on Large Language Models." https://arxiv.org/abs/2411.00154

• results.zip

To access the results, unzip the file results.zip (link).

You will see folders for each experiment setup (i.e., collection, document, sentence, continual training, and fine-tuning). Inside each folder, you will see the results organized by model. We did experiments on Pythia 2.8B, Pythia 6.9B, and GPT Neo 2.7B.

The main files we include are:

• The precomputed MIA attacks are stored in results/{data_scale}/EleutherAI/{model}/haritzpuerto/{data_partition}/mia_members.jsonl and mia_nonmembers.jsonl
• The CSV files with the evaluation performance are stored in results/{data_scale}/EleutherAI/{model}/haritzpuerto/{data_partition}/*.csv
• For each data partition, the used to conduct the experiments. They are stored in results/{data_scale}/EleutherAI/{model}/haritzpuerto/{data_partition}/members and non_members. You need to open them with datasets.load_from_disk

The precomputed MIA attacks are stored as a list of jsons. Each json has the following form:

Extract from results/collection_mia/EleutherAI/pythia-6.9b/haritzpuerto/the_pile_00_arxiv/2048/mia_members.jsonl

{
   "pred":{
      "ppl":9.5,
      "ppl/lowercase_ppl":-1.028301890685848,
      "ppl/zlib":0.00022461257094747036,
      "Min_5.0% Prob":9.479779411764707,
      "Min_10.0% Prob":8.171262254901961,
      "Min_20.0% Prob":6.549893031784841,
      "Min_30.0% Prob":5.498956636807818,
      "Min_40.0% Prob":4.719867435819071,
      "Min_50.0% Prob":4.099095796676441,
      "Min_60.0% Prob":3.588011502442997
   },
   "label":1
}

The csv results are tables like the following:

Extract from results/collection_mia/EleutherAI/pythia-6.9b/haritzpuerto/the_pile_00_arxiv/2048/dataset_inference_pvalues_10_dataset_size.csv

Dataset Size	Known Datasets	Training Size	Eval Size	F1	P-value	TPR	FPR	AUC	Chunk-level AUC	Seed
10	1000	2000	2000	57.07246213473086	0.4321467209427013	52.900000000000006	38.6	0.593152	0.5275510595912055	670487
10	1000	2000	2000	56.79208146268461	0.555579505655733	70.3	55.300000000000004	0.5959169999999999	0.5277849316855144	116739

Please refer to our 2025 NAACL Findings paper "Scaling Up Membership Inference: When and How Attacks Succeed on Large Language Models" for all the details to understand and interpret the results.

Developed at Parameter Lab with the support of Naver AI Lab.

Disclaimer

This repository contains experimental software results and is published for the sole purpose of giving additional background details on the respective publication.

Citation

If this work is useful for you, please consider citing it

@misc{puerto2024scalingmembershipinferenceattacks,
      title={Scaling Up Membership Inference: When and How Attacks Succeed on Large Language Models}, 
      author={Haritz Puerto and Martin Gubri and Sangdoo Yun and Seong Joon Oh},
      year={2024},
      eprint={2411.00154},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2411.00154}, 
}

✉️ Contact person: Haritz Puerto, [email protected]

🏢 https://www.parameterlab.de/

🌐 https://haritzpuerto.github.io/scaling-mia/

RT.AI https://researchtrend.ai/papers/2411.00154

Don't hesitate to send us an e-mail or report an issue if something is broken (and it shouldn't be) or if you have further questions.