Extracting Prompts from Customized Large Language Models

This paper consists of the source code of paper: Why Are My Prompts Leaked? Unraveling Prompt Extraction Threats in Customized Large Language Models(arxiv).

Source code explanations

• PEAD dataset: extractingPrompt/instructions/benchmark_collections/OVERALL_DATA_BENCHMARK.json
• Source code of all experiments: extractingPrompt/
  • Generalized evaluation
    • Vanilla: extractingPrompt/1.run_prompt_extraction.py
    • Function callings comparison: extractingPrompt/5.funcall_comparison.py
  • Scaling laws of prompt extraction
    • Model size: extractingPrompt/2.model_size_prompt_extraction_experiments.py
    • Sequence length: extractingPrompt/4.varying_sequence_length.py
  • Empirical analysis
    • Convincing Premise: extractingPrompt/6.ppl_comparison.py
    • Parallel-translation: extractingPrompt/7.attention_visualize.py
    • Parallel-translation: extractingPrompt/attention_visualize.py
  • Defense strategies
    • Defending methods: extractingPrompt/defending/ppl_high2_confusingBeginnings.py
    • Performance drops experiments of the defending: extractingPrompt/defending/2.drops_of_defending.py
    • visualization: extractingPrompt/defending/defense_visualization.py
  • Close-AI experiments
    • vanilla prompt extraction: extractingPrompt/api_related_experiments/1.run_prompt_extraction.py
    • soft extraction: extractingPrompt/api_related_experiments/2.soft_extraction_experiments.py
    • performance drops of defending: extractingPrompt/api_related_experiments/3.1.drops_of_defense.py

Experimental environments

Run

pip install -r re.txt

or install the following key packages manually:

datasets
numpy
pandas
peft
safetensors
scipy
tensorboard
tensorboardX
tiktoken
tokenizers
torch
tqdm
transformers
matplotlib
scikit-learn
thefuzz
einops
sentencepiece

Contact the authors

Feel free to open an issue, or send the email to zi1415926.liang@connect.polyu.hk if there exists any problem.

Citation:

@misc{liang2024promptsleakedunravelingprompt,
      title={Why Are My Prompts Leaked? Unraveling Prompt Extraction Threats in Customized Large Language Models}, 
      author={Zi Liang and Haibo Hu and Qingqing Ye and Yaxin Xiao and Haoyang Li},
      year={2024},
      eprint={2408.02416},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2408.02416}, 
}