HotSwapPII

A comprehensive platform for end-to-end evalution of PII detection and anonymizing in unstrcutured text documents using various SOTA models.

Home Page	Results View

Overview

This application provides an interactive interface for detecting PII entities in text with features including:

• Support for multiple NER engines (SpaCy, HuggingFace Transformers, GLiNER)
• Multiple anonymization methods (redaction, replacement, masking, encryption)
• Synthetic data generation with OpenAI
• Model evaluation capabilities
• Entity-type performance metrics
• Customizable detection settings

Installation

Using Poetry (Recommended)

1. Clone this repository
2. Create a virtual environment
3. Install dependencies with Poetry:

# Install Poetry if needed
pip install poetry

# Install dependencies
poetry install

# Run the application
poetry run streamlit run app.py

Using pip

1. Clone this repository
2. Install dependencies:

pip install -r requirements.txt

# Run the application
streamlit run app.py

Usage

PII Detection

1. Enter or upload text in the input panel
2. Configure model and detection settings in the sidebar
3. Click "Analyze Text" to detect PII entities
4. View detection results and anonymized output

Model Evaluation

1. Navigate to the "Model Evaluation" tab
2. Upload a CSV file with labeled data
3. Configure evaluation settings
4. Click "Evaluate Model" to assess model performance
5. View evaluation metrics and entity-type performance

Features

Detection Models

• SpaCy: Fast and efficient NER models for production use
• HuggingFace Transformers: Deep learning models with higher accuracy
• GLiNER: Generalist model for zero-shot named entity recognition

Anonymization Methods

• Redact: Remove PII entities completely
• Replace: Replace with entity type (e.g., <PERSON>)
• Mask: Replace characters with a mask character (e.g., *******)
• Hash: Replace with a hash of the text
• Encrypt: Encrypt the text (reversible)
• Highlight: View original text with highlighted entities
• Synthesize: Replace with realistic fake values using OpenAI

Advanced Features

• Allow/Deny Lists: Customize detection with word lists
• Custom Regex Patterns: Define your own entity patterns
• Overlap Handling: Configurable handling of overlapping entities
• Entity Selection: Choose which entity types to detect
• Decision Process: View reasoning behind detection decisions

Data Format for Evaluation

The evaluation feature expects a CSV file with:

• A text column containing the text to evaluate
• A label column with JSON annotations in this format:

[{"start": 58, "end": 89, "text": "John Smith", "labels": ["name"]}]

Project Structure

pii-detection/
├── app.py                  # Main application entry point
├── config/                 # Configuration settings
├── models/                 # Model implementations
├── core/                   # Core detection and anonymization logic
├── ui/                     # UI components
├── utils/                  # Utility functions
├── data/                   # Sample data and resources
└── README.md               # Project documentation

Dependencies

• presidio-analyzer: Microsoft's PII detection library
• presidio-anonymizer: Microsoft's PII anonymization library
• streamlit: Web application framework
• spacy: NLP toolkit
• transformers: Hugging Face's transformer models
• pandas: Data manipulation
• openai: OpenAI API for synthetic data

Credits

• This application uses Microsoft Presidio for PII detection and anonymization
• UI built with Streamlit