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.

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