PrivacyShield

Redact sensitive PDFs locally, share safely, restore anytime with your personal key.

PrivacyShield is a local-first PDF redaction system that automatically detects and blacks out personal information (PII) in PDF documents with no data ever leaving your device. Every redaction is reversible: a unique encryption key lets the document owner restore original content at any time.

---

Features

• Automatic PII Detection — names, SSNs, IBANs, phone numbers, emails, addresses, IDs, medical conditions, financial amounts, and more
• Multilingual — English, German, French, Italian, Spanish
• Text + Scanned PDFs — pdfplumber for text-based pages, pypdfium2 + PaddleOCR for scanned/image pages
• Reversible Redaction — encrypted .privacyshield key file lets you restore original values
• 100% Local — no cloud, no API calls, no data leaves your machine
• Simple Web UI — drag-and-drop PDF upload, one-click download
• REST API — FastAPI backend with /redact and /unredact endpoints for programmatic access
• Swiss PII Support — AHV/AVS numbers, IBAN validation, RF creditor references

---

How It Works

PDF Input
    ↓
Analyzer      →  classify each page: text / scanned / mixed
    ↓
Text pages    →  Extractor → NER Engine → Redactor → PDF Rebuilder
Scanned pages →  pypdfium2 → PaddleOCR → Image Redactor
Mixed pages   →  both pipelines run and results are merged
    ↓
Key Manager   →  encrypt token map → .privacyshield file (Fernet AES-128)
    ↓
Output: Redacted PDF  +  Encryption Key (shown once to user)

To restore: upload redacted PDF + paste your key → original values decrypted and restored.

---

Detected PII Types

Category	Examples
Person names	Full names, partial names
Contact info	Email, phone (international formats)
National IDs	SSN, passport, driver's license
Swiss-specific	AHV/AVS number (756.XXXX.XXXX.XX)
Financial	IBAN (with mod-97 validation), SWIFT/BIC, salary amounts
Document IDs	Policy numbers, invoice numbers, claim numbers, UUIDs, TAX IDs, RF references
Medical	Diagnosis labels, condition names
Location	Addresses

---

Project Structure

privacyshield/
├── app.py                          ← Flask web application
├── streamlit_app.py                ← Streamlit UI (alternative)
├── requirements.txt
├── privacyshield/
│   ├── analyzer/
│   │   └── pdf_analyzer.py         ← Classify pages: text/scanned/mixed
│   ├── text_pipeline/
│   │   ├── extractor.py            ← Extract text + char coordinates
│   │   ├── ner_engine.py           ← PII detection (Presidio + spaCy)
│   │   ├── redactor.py             ← Token replacement
│   │   └── pdf_rebuilder.py        ← Draw black boxes (PyMuPDF)
│   ├── image_pipeline/
│   │   ├── pdf_to_image.py         ← Convert PDF page → PIL Image (pypdfium2)
│   │   ├── ocr_engine.py           ← PaddleOCR text + coordinates
│   │   ├── image_classifier.py     ← Classify image regions (photo/scanned/id_card)
│   │   └── image_redactor.py       ← Draw boxes on image layer with token labels
│   ├── key_manager/
│   │   ├── encryptor.py            ← Fernet encryption
│   │   └── decryptor.py            ← Fernet decryption
│   ├── reconstructor/
│   │   └── pdf_merger.py           ← Merge text + image redactions into final PDF
│   ├── templates/
│   │   └── index.html              ← Single-page web UI
│   └── pipeline.py                 ← Orchestrates full pipeline
├── api/
│   ├── main.py                     ← FastAPI app
│   ├── routes/
│   │   ├── redact.py               ← POST /redact endpoint
│   │   ├── unredact.py             ← POST /unredact endpoint
│   │   └── health.py               ← GET /health endpoint
│   └── models/
│       └── schemas.py              ← Pydantic request/response models
└── testing/
    └── GSF/                        ← 100 synthetic test documents

---

Prerequisites

• Python 3.11
• pip
• No system dependencies required i.e. pypdfium2 bundles its own PDF renderer and works on Windows, macOS, and Linux without poppler

Note: First run will download spaCy language models (~2GB total) and PaddleOCR models (~200MB) automatically.

---

Installation and Run

1. Clone the repository

git clone https://github.com/DebDDash/privacyshield.git
cd privacyshield

2. Create a virtual environment

macOS / Linux:

python3 -m venv venv
source venv/bin/activate

Windows (PowerShell):

python -m venv venv
venv\Scripts\Activate.ps1

If PowerShell blocks the activation script, run this once first:

Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass

3. Install dependencies

pip install -r requirements.txt

Total install size is approximately 2.5–3GB due to PaddleOCR and spaCy models.

4. Run the Flask web app

python app.py

5. Open in browser

http://127.0.0.1:5000

6. Optional — Run FastAPI backend instead

uvicorn api.main:app --reload --port 8000

Then open the interactive API docs at:

http://127.0.0.1:8000/docs

---

Usage

Redact a PDF

1. Open http://127.0.0.1:5000
2. Drag and drop your PDF onto the upload area (max 50MB)
3. Click Upload & Process
4. Wait for processing (30–120 seconds depending on PDF size)
5. Copy and save your Recovery Key — shown once, never stored
6. Click Download Redacted PDF

Restore Original Values

1. Click the Decrypt PDF tab
2. Upload the redacted PDF
3. Paste your Recovery Key
4. Click Restore Original PDF
5. Download the restored document

---

Security Model

Original PDF  ──→  Redaction  ──→  Redacted PDF (safe to share)
                       │
                       └──→  Token Map  ──→  Fernet Encrypt  ──→  .privacyshield
                                                   │
                                             Recovery Key
                                          (shown once to user,
                                           never stored on disk)

• The .privacyshield file contains the encrypted mapping of tokens to original values
• The Recovery Key uses Fernet (AES-128-CBC + HMAC-SHA256)
• Neither the key nor the original values are ever stored by the application
• The redacted PDF alone reveals nothing — you need both the redacted PDF and the key to restore

---

Supported Languages

Language	spaCy Model	PII Detection
English	en_core_web_lg	Full
German	de_core_news_lg	Full + AHV/AVS
French	fr_core_news_lg	Full + AVS
Italian	it_core_news_lg	Full
Spanish	es_core_news_lg	Full

Language is auto-detected per page using langdetect.

---

Troubleshooting

Port 5000 already in use:

# macOS/Linux
lsof -i :5000
kill -9 <PID>
python app.py

spaCy model not found:

python -m spacy download en_core_web_lg
python -m spacy download de_core_news_lg
python -m spacy download fr_core_news_lg
python -m spacy download it_core_news_lg
python -m spacy download es_core_news_lg

PaddleOCR slow on first run: First run downloads OCR models (~200MB). Subsequent runs use cached models and are significantly faster.

numpy ABI error on startup:

RuntimeError: module compiled against ABI version 0x1000009

Fix with:

pip install "numpy<2.0"

PDF processing fails:

• Ensure the PDF is not password-protected
• Check the terminal for detailed error messages

---

Team

Built for the GenAI Zürich Hackathon 2026 — GoCalma Privacy Redaction Track. Members:

1. Debarpita Dash
2. Shrimi Agrawal
3. Sruthi Subramanian
4. Pragati Agrawal

---

License

MIT License — see LICENSE for details.