Privately

Privately

Privately is a Chrome extension that protects users from unintentionally leaking sensitive information in prompts before submitting them to AI systems. It is designed for developers, students, and knowledge workers who frequently include code, configs, or personal data into AI tools.

Our tool is specifically catered to target PII within code contexts and especially in the context of Singapore.

Privately combines regex-based validators (for exact identifiers like NRIC, phone numbers, credit cards) with a hosted FastAPI inference server running a fine-tuned DistilBERT-base-uncased model to catch fuzzy PII such as names and addresses.

---

🔄 How It Works

1. User Input

   • User enters prompts into a AI tool as usual. Prompts can include natural language and code.
   • Chrome extension intercepts the input before submission and provide annotations to inputted prompt.

2. Local Regex Validation

   • Immediate scanning for structured PII (e.g. phone numbers, NRIC, credit cards).

3. Server Model Inference (FastAPI + ONNX Runtime)

   • For unstructured / fuzzy entities (names, addresses), the extension sends the text to a FastAPI backend.
   • Backend loads the fine-tuned DistilBERT model (quantized to ONNX INT8 for performance).
   • API responds with entity spans and labels.

4. UI Feedback

   • Detected entities are underlined in the browser.
   • Inline tooltips allow:
     • Redaction
     • Replacement with custom placeholders
     • Skip current
     • Ignore all

5. Submission

   • After review/cleanup, the user submits the text to the AI system without leaking sensitive data.

---

✨ Features

• Smart PII Detection
  Combines fast regex validators (for structured data like emails, phone numbers, NRIC, credit cards) with a fine-tuned DistilBERT-base-uncased model (for fuzzy PII like names and addresses). Detection runs in real-time on every input box in Chrome.

• Inline Tooltips

  • Display detected text and display category (e.g., NAME, ADDR, EMAIL).
  • One-click actions: Remove, Replace with custom placeholder, Skip once, or Ignore all for that category.

• User Control & Transparency

  • Categories are fully toggleable in the options panel.
  • Tooltip always shows both the detected text and its probable category, so you understand why it was flagged.

• Seamless Chrome Integration

  • Works across all websites in the browser.
  • Lightweight extension with no external dependencies.
  • All detections happen before submission

---

🗂️ Detection Categories

✅ Regex-Based Categories

• Emails (standard + RFC validation)
• Phone Numbers
• URLs & IP Addresses
• NRIC/FIN
• Postal Codes
• Credit Cards
• JWTs
• Authorization Headers, Cookies
• File Paths
• UUID/GUIDs

🤖 Model-Based Categories (via DistilBERT-base-uncased + ONNX)

• PER (Names)
  • Multicultural names (Chinese, Malay, Indian, Western, initials, hyphenated).
• ADDR (Addresses)
  • Global + Singapore addresses (HDB, condo, commercial, postal, block/unit).

---

Fine-Tuning DistilBERT-base-uncased

📊 Dataset Creation

Since real-world PII datasets within code prompts are difficult to obtain, synthetic dataset is created to train the model

1. Source Data

   • CSV file (pii_database.csv) with name and address columns.
   • Contains multicultural names (Malay, Chinese, Indian, Western) and Singapore-style addresses (HDB blocks, streets, condos, etc.).
   • Name and Address data is obtained from Singapore based datasets obtainable online.

2. Synthetic Sentence Generation

   • A Python script reads from the CSV and generates natural + code-like snippets.
   • Embeds PII in code like environment
   • Adds Fuzz and variation

3. Auto-Annotation

   • Every generated text snippet includes offsets (start, end) pointing to the NAME and ADDR spans.

4. Export Format

   • Final dataset is stored in JSONL format for training.

---

🧑‍🏫 Model Training

The model is a fine-tuned DistilBERT-base-uncased for token classification:

1. Preprocessing

   • Tokenized text with AutoTokenizer.from_pretrained("distilbert-base-uncased").
   • Converted annotated spans into BIO tags:
     • B-NAME, I-NAME for names
     • B-ADDR, I-ADDR for addresses
     • O for non-PII tokens

2. Fine-Tuning

   • Used Hugging Face Trainer API:
   • Trained for multiple epochs with cross-entropy loss on token labels.
   • Split dataset into train / validation (e.g., 80/20).

3. Saving & Export

   • Saved the Hugging Face model:
   • Exported to ONNX
   • Quantized to INT8 for faster inference

4. Deployment

   • The quantized ONNX model (onnx_int8/) is served with FastAPI.
   • Extension sends text → FastAPI → ONNX Runtime inference → returns detected PII spans.

---

📈 Why This Approach?

• Robustness: Fuzzing (case, spacing, typos, noise) improves generalization.
• Efficiency: ONNX + INT8 quantization reduces model size and speeds up inference.
• Scalability: Hosted via FastAPI → can be deployed serverless or containerized.

🛠️ Tech Stack

• Frontend:

  • Chrome Extension (Manifest V3, content scripts + service worker).
  • Regex validator in JavaScript

• Backend (PII Detection):

  • Fine-tuned DistilBERT-base-uncased → exported to ONNX (INT8).
  • FastAPI serving inference with ONNX Runtime.

• Training:

  • Fine-tuned on Singapore-specific datasets (names, addresses, orgs).
  • Hugging Face Transformers → Optimum ONNX → quantized INT8.

• API Communication:

  • Chrome extension → FastAPI /detect endpoint.
  • Hosted on Render
  • JSON response with spans + labels.

---

🚀 Installation

1. Clone this repo:

   git clone https://github.com/your-org/privately.git
   cd privately
   

2. Load extension directory into chrome browser

Project Structure

privately/
├─ extension/                 # Chrome extension (MV3)
│  ├─ manifest.json
│  ├─ assets/
│  └─ src/
│     ├─ content.js           # inline detection + tooltip
│     ├─ overlay.css          # styles for highlights/tooltips
│     ├─ popup.html
│     ├─ popup.js             # dashboard popup
│     ├─ options.html
│     └─ options.js           # settings (categories, modes)
├─ server/                    # FastAPI backend for PII model
│  ├─ main.py                 # FastAPI app (model inference)
│  ├─ onnx_int8/              # Model + tokenizer files
├─ web-dashboard/             # Lynx web dashboard
└─ README.md