DataShield

Inspiration

As AI agents like ChatGPT, Claude, Gemini, and Copilot become embedded in daily workflows, users increasingly paste sensitive data, SSNs, addresses, emails, internal company information, into prompt boxes without realizing the long-term implications.

We faced a core dilemma: Productivity vs. Privacy

Existing redaction tools often send data to another server for processing-which defeats the purpose of privacy.

We wanted a solution that:

Runs 100% locally
Requires zero trust in third-party servers
Integrates seamlessly into AI workflows

That vision became DataShield.

What it does

DataShield is a Chrome Extension that acts as an intelligent privacy layer for ChatGPT, Gemini, Copilot, and Claude.

When a user types into an AI prompt box, DataShield provides:

🔎 Total Risk Score: An aggregated risk metric computed from detected PII entities. 📋 Granular Detection: A categorized breakdown of all detected Personally Identifiable Information, including risk level per entity type. 🎛 Selective Masking: Users can choose exactly which detected data points to redact.

With one click, sensitive text is replaced with masked tokens-before the data leaves the browser.

How we built it

We followed a structured three-stage architecture:

1️⃣ Fine-Tuning the Detection Model To keep inference lightweight and browser-compatible, we chose DistilBERT (distilbert-base-cased) and fine-tuned it for Named Entity Recognition (NER).

Dataset: ai4privacy/pii-masking-200k Subset size: ~20,000 high-quality samples GPU: T4 on Google Colab Target metric: F1 > 0.9

The result was a 54-class PII detection model optimized for contextual understanding.

2️⃣ Extension Development We developed a Manifest V3 Chrome extension featuring:

Robust content scripts
Dynamic interaction with modern AI prompt text areas
A popup UI with real-time entity classification and risk visualization

3️⃣ ONNX-Based Browser Inference To run inference locally in-browser:

We converted the trained model to ONNX
Used in-browser execution via ONNX runtime
Implemented IOB (Inside-Outside-Beginning) tagging logic
Combined regex + model inference: Transformer → unstructured data, Regex → structured data (e.g., SSNs, phone numbers)

We also built custom token-stitching logic to resolve subword splits (e.g., App + ##le → Apple) without breaking PII classification.