ShuzeLiu11/mmd-llm-mea-detection-data

MMD LLM MEA Detection Data

This dataset repository contains the processed query files used by the paper:

An Embarrassingly Simple Detector for Model Extraction Attacks in Large Language Model API Traffic

Paper: https://arxiv.org/abs/2606.05725
Code: https://github.com/LabRAI/mmd-llm-mea-detection

Dataset Description

The dataset is a processed research artifact for evaluating model extraction attack detection in LLM/API query traffic. Each attacker-normal pair contains two query pools:

data/<dataset>/normal/queries.jsonl
data/<dataset>/attacker/queries.jsonl

The normal query pool is used to construct benign reference/calibration windows and benign evaluation windows. The attacker query pool is used to construct pure attacker windows and mixed benign-attacker traffic windows.

Data Format

Each file is stored in JSONL format. Each row contains a text query under the query field:

{"id": "example-0", "index": 0, "query": "What is the capital of France?"}

Usage

To use this dataset with the companion code repository:

git clone https://github.com/LabRAI/mmd-llm-mea-detection.git
cd mmd-llm-mea-detection

hf download ShuzeLiu11/mmd-llm-mea-detection-data \
  --repo-type dataset \
  --local-dir data

Then run an experiment, for example:

DATASET=model_leeching \
BATCH_SIZE=1500 \
DEVICE=cuda \
MMD_DEVICE=cuda \
SEED=42 \
bash MMD_detection/run_mmd_detection_gpu.sh

Source and License Notes

This dataset is a mixed-source processed research artifact derived from public datasets and model-extraction query generation procedures. The original data sources retain their respective licenses and terms of use. Users are responsible for complying with the licenses and terms of the underlying source datasets, including SQuAD, GLUE, BoolQ, AG News, SST-2, IMDB, WikiText-103, WildChat, and related model extraction attack artifacts where applicable.

The processed files are provided for research and reproducibility purposes. They should not be used to identify individuals, violate source dataset terms, or deploy harmful model extraction attacks.

Citation

If you use this dataset, please cite:

@misc{liu2026embarrassinglysimpledetectormodel,
      title={An Embarrassingly Simple Detector for Model Extraction Attacks in Large Language Model API Traffic}, 
      author={Shuze Liu and Qianwen Guo and Yushun Dong},
      year={2026},
      eprint={2606.05725},
      archivePrefix={arXiv},
      primaryClass={cs.CR},
      url={https://arxiv.org/abs/2606.05725}, 
}