This repository contains the source code for SaMIA: Sampling-based Pseudo-Likelihood for Membership Inference Attacks.

WikiMIA, a dataset for evaluating MIA, has been placed under the directory wikimia/.

Samples can be generated using the following command:

python src/sampling.py --model_name gpt-j-6B --text_length 32 --num_samples 10 --prefix_ratio 0.5

where

• model_name specifies the model used for sampling, chosen from {gpt-j-6B, opt-6.7b, pythia-6.9b, Llama-2-7b`};
• text_length specifies the length of texts, chosen from {32, 64, 128, 256};
• num_samples specifies the number of samples to be generated;
• prefix_ratio specifies the percentage of texts used as the prefix (prompt for generating the following text).

The generated samples will be collected in sample/{model_name}/{text_length}.jsonl.

!Note that in order to use Llama-2-7B, be sure to obtain the permission for access and assign your huggingface token to variable YOUR_HUGGINGFACE_TOKEN.

We evalaute the leakage detection performance of SaMIA based on the surface similarity between generated samples, under the directory sample/, and the original texts, under the directory wikimia/. The surface similarity is measured using ROUGE-N (N=1). Evaluation can be conducted using the following command:

python src/eval_samia.py  --model_name gpt-j-6B --text_length 32 --num_samples 10 --prefix_ratio 0.5

Argument model_name, text_length, num_samples, and prefix_ratio functions similarly as in the previous section, with two additional flags available here:

• --zlib enables zlib entropy compression;
• --save dumps the computed rouge-n results of leaked (seen) and unleaked (unseen) texts into files.

The source code will print AUC-ROC and TPR@10%FPR at stdout.

If you find this repository useful for your research, please cite us with:

@misc{kaneko2024samplingbased,
      title={Sampling-based Pseudo-Likelihood for Membership Inference Attacks}, 
      author={Masahiro Kaneko and Youmi Ma and Yuki Wata and Naoaki Okazaki},
      year={2024},
      eprint={2404.11262},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}