A5 PII Anonymizer

Built-In LLM Desktop App Preview:

This repository provides an Electron desktop application for locally anonymizing documents before sending them to advanced Large Language Models (LLMs). By stripping out personal or identifiable information using a context-aware model, you can safely train or query external LLMs (e.g., OpenAI’s o3 model) with minimal privacy risk.

Motivation

• PII Removal: Traditional RegEx-based anonymization often fails on nuanced data. With an ONNX-based model, you gain context-aware detection for names, addresses, phone numbers, etc.
• Safe LLM Usage: Many companies need to keep real customer or employee data internal but still want to leverage powerful external LLMs. This tool helps them do so by anonymizing data on their end first.
• Flexible: Supports .txt, .docx, .xls(x), .csv, .pdf, and more. Converts text, merges tokens, and replaces them with consistent pseudonyms.

Key Features

1. Electron App: Cross-platform desktop UI built on HTML/CSS/JS.
2. Daily Limit: 100 documents per day for the free tier (by default). You can raise or remove it if you prefer—this is open source.
3. Context-Aware: Relies on a local ONNX model downloaded separately (due to GitHub’s file-size constraints).
4. Mapping (Pro Mode): If you enable Pro, the app can produce a JSON file mapping each original entity (e.g., “John Smith”) to its anonymized token (e.g., “NAME_1”).
5. MIT License: Free to modify and distribute. We welcome contributions.

Getting Started

1. Clone or Download this repository.
   • Model available here: https://huggingface.co/iiiorg/piiranha-v1-detect-personal-information/tree/main
   • Download the ONNX model (.onnx files) from our external link (not included here due to size constraints).
   • Place it under ./models/protectai/lakshyakh93-deberta_finetuned_pii-onnx/ or as directed in fileProcessor.js.
2. Install Dependencies:

   npm install

3. Run (Dev Mode):

   npm run dev

   Or if you prefer:

   npx electron .

4. Build / Package (macOS example):

   npm run build:mac
   
   

Basic Usage

• Drop or Select Files: The main UI allows you to drag-and-drop or pick multiple files/folders.
• Output Directory: Choose where the anonymized files should be placed.
• Anonymize: Click “Anonymize Files” to run.
• Mapping (Pro): If you have a Pro key, the app will create an additional -map.json file capturing each replaced entity.

How It Works

• Electron:
  • main.js: Spawns the main window, handles file selection, passes tasks to FileProcessor.
  • renderer.js: Manages the UI (index.html), user interactions, daily usage counters, and “Pro” logic.
• fileProcessor.js:
  • Loads the local ONNX model (via @xenova/transformers).
  • Identifies personal data by context (names, addresses, etc.).
  • Replaces them with tokens (NAME_1, PHONE_NUMBER_3, etc.).
  • If Pro, writes a JSON mapping for re-identification.
• Local Model:
  • We rely on a context-aware token classification model. This is significantly more effective than simple RegEx for real-world PII.

Limitations & Notes

• Daily 100-File Limit: By default, the free version only processes 100 documents per day. This is purely enforced in the UI. Since it’s open source, you can remove or change it as needed.
• No Guarantee: Even context-aware models can miss certain edge cases. Always manually review if 100% privacy is critical.
• Cross-Platform:
  • macOS builds are tested on both M-series (ARM) and Intel.
  • Windows and Linux builds are also available.

Contributing

We use the MIT License, so feel free to open pull requests, modify the code, or adapt it for your needs. If you make improvements or fix bugs, please share them back!

---

Thanks for checking out the A5 PII Anonymizer. We hope this helps you safely leverage powerful LLMs with real data while keeping personal information private.