Code for Rap-LI

🛠️ Installation

1. Set up the Rap_part1 environment

conda create -n Rap_part1 python=3.10.15
conda activate Rap_part1
pip install -r requirements_part1.txt

2. Set up the Rap_part2 environment

This environment is exclusively used for LLM inference to avoid package conflicts.

conda create -n Rap_part2 python=3.10.15
conda activate Rap_part2
pip install -r requirements_part2.txt

3. Note on Specific Folders

💡 Note: When running code in specific subfolders (e.g., interactive_demo), please check their respective README.md and requirements.txt for additional environment requirements.

💡 Troubleshooting: If encountering package conflicts, manually install dependencies using:

pip install <package_name>==<specific_version>

---

🚀 Usage

Phase 0: Data Preparation

Prepare the datasets and embeddings for data/ folder.

Phase 1: Prompt Sanitization and Privacy Evaluation

🔑 Activate the Rap_part1 environment and run the following command:

conda activate Rap_part1
python _run_script.py

🤖 The script will batch process all methods on all downstream tasks, including privacy risk identification, prompt sanitization, attack evaluation, and privacy analysis.

🏷️ To run a specific method, you can run "main_...py" or "baseline_...py" in the root directory or in the baseline folder. Don't forget to modify the args file to configure parameters (such as tasks or datasets). 🔖 You can run the files in Privacy_Analysis folder to get the privacy analysis results.

🔖 You can run the files in intermediate_metrics_Analysis folder to get the Intermediate Metrics that measure text quality after sanitization.

Phase 2: LLM Inference

🔨 update the main_blackllm.py file by replacing the black_llm_model_name and api_key with your own values. Ensure that the base_url in _04black_llm\black_llm_interface.py is correct.

🔑 Activate the Rap_part2 environment and run the following command:

conda activate Rap_part2
python main_blackllm.py

🤖 It will run LLM inference for all methods on all downstream tasks.

---

📦 Finally, you can get detailed results and brief reports with json and csv files.

---

💻 Interactive Demo: Privacy-Preserving LLM Chat System

To validate the practical applicability and usability of Rap-LI, we designed a Privacy-Preserving LLM Chat System (similar to ChatGPT or Gemini) that integrates the proposed Rap-LI with an interactive interface.

Key Features

• Interactive Risk Adjustment: Users can manually adjust token risk levels via an interactive pop-up window. The system automatically marks privacy entities based on Rap-LI's defaults, which users can then fine-tune.
• Side-by-Side Comparison: To enhance user experience, the system displays the original prompt and the sanitized prompt side-by-side. Replaced entities are highlighted using colors corresponding to their risk levels.
• One-Click Restoration: The system provides a functional module to extract and restore original private entities from the LLM output, improving readability and task completion.
• Customizable Sensitivity Database: Supports industry- or user-specific sensitivity databases. Custom rules are automatically mapped to risk levels to reduce manual adjustments.

Resources

• Source Code: The complete system code is available in the interactive_demo/ folder. Please refer to interactive_demo/README.md for detailed setup and usage instructions.
• Video Tutorial & Demo: You can watch the system introduction, usage tutorial, and recorded user study videos here: Google Drive Link

---