AntiWatermark

Local-only AI text watermark remover, steganography stripper, and humanizer.

Strips invisible Unicode markers, replaces AI clichés, validates output integrity, and optionally rewrites text via a local AI model. No cloud services. No API keys. No data leaves your machine.

🔒 100% Local — No Data Leaves Your Machine

---

Quick Start

git clone https://github.com/policeakshithreddy/antiwatermark
cd antiwatermark

python -m venv .venv
source .venv/bin/activate

pip install -e .

pytest

python app.py

Open http://localhost:8000

This works without any AI model. Fast Mode is always available.

---

Two Modes

Fast Clean (Default)

Deterministic, instant, zero dependencies.

• Strips 26+ invisible Unicode characters (zero-width spaces, BOM, soft hyphens, etc.)
• Replaces 30+ AI clichés ("delve into", "rich tapestry", "multifaceted", "Certainly!")
• Protects code blocks, inline code, LaTeX, URLs, Markdown links, and emails
• Validates output integrity
• Runs heuristic diagnostics

Local Neural (Optional)

Uses a local AI model via Ollama for deeper text rewriting.

# Install Ollama (https://ollama.com)
ollama pull llama3.2
ollama serve

# Then select "Local Neural" in the UI

No cloud APIs. No API keys. Ollama runs entirely on your machine.

---

Installation

Base (Fast Mode only):

pip install -e .

With neural support:

pip install -e ".[neural]"

---

Usage

Web UI

python app.py
# Open http://localhost:8000

CLI

# Clean text
antiwatermark "Your AI-generated text here"

# Clean a file in-place
antiwatermark document.txt --inplace

# Output diagnostics as JSON
antiwatermark "text" --json

# Use Ollama backend
antiwatermark "text" --backend ollama

# Launch Web UI
antiwatermark --web

# Run clipboard daemon
antiwatermark --daemon

Python API

from antiwatermark import clean_text, validate_output, ImmunityShield

# Fast clean
cleaned, diagnostics = clean_text("Your AI text here")
print(cleaned)
print(f"Human confidence: {diagnostics['human_confidence_pct']}%")

# With validation
shield = ImmunityShield()
shield.shield(original_text)
result = validate_output(original_text, processed_text, shield)
if result.is_valid:
    print("Output validated successfully")
else:
    print(f"Validation failed: {result.failures}")

REST API

# Health check
curl http://localhost:8000/api/health

# Fast clean
curl -X POST http://localhost:8000/api/clean \
  -H 'Content-Type: application/json' \
  -d '{"text": "Your text here"}'

# Neural rewrite (requires Ollama)
curl -X POST http://localhost:8000/api/rewrite \
  -H 'Content-Type: application/json' \
  -d '{"text": "Your text here", "model": "llama3.2"}'

# Configuration
curl http://localhost:8000/api/config

---

Chrome Extension

1. Open chrome://extensions
2. Enable Developer mode
3. Click Load unpacked → select the extension/ directory
4. The extension automatically strips invisible characters when copying
5. Right-click selected text → AntiWatermark → Clean Selection or Rewrite Selection Locally

---

Project Structure

AntiWatermark/
├── antiwatermark/
│   ├── core.py          # Cleaning, shielding, validation, diagnostics
│   ├── backend.py       # RewriterBackend, BuiltinBackend, OllamaBackend
│   ├── middleware.py     # Rewrite orchestration, domain personas
│   ├── cli.py           # Command-line interface
│   └── __init__.py
├── web/
│   └── index.html       # Self-contained Web UI
├── extension/
│   ├── manifest.json
│   ├── background.js
│   ├── content.js
│   ├── popup.html
│   └── popup.js
├── tests/
│   ├── test_core.py
│   ├── test_backend.py
│   └── test_api.py
├── scripts/
│   └── clipboard_daemon.py
├── app.py               # Local HTTP server
├── pyproject.toml
└── README.md

---

Testing

pytest

82+ tests covering:

• Invisible Unicode removal
• Cleaning idempotency
• Code/LaTeX/URL/Markdown/email preservation
• Collision-safe placeholders
• Validation gates
• Builtin backend
• Local-only security policy
• API endpoints

---

What AntiWatermark Does

Feature	Description
Strip invisible chars	Removes 26+ zero-width Unicode characters used for text fingerprinting
Replace AI clichés	Deterministically replaces 30+ common AI writing patterns
Protect structured content	Shields code blocks, LaTeX, URLs, links, and emails from modification
Validate output	5-gate validation ensures protected content is preserved
Heuristic diagnostics	Analyzes text patterns, burstiness, and AI marker density
Local neural rewrite	Optional deeper rewriting via local Ollama models

---

What AntiWatermark Does NOT Do

• ❌ No cloud API calls
• ❌ No data exfiltration
• ❌ No API keys required
• ❌ No torch/transformers dependencies
• ❌ No automatic model downloads
• ❌ No detector score manipulation claims

---

Security

All neural communication is restricted to loopback addresses only:

• 127.0.0.1 ✅
• localhost ✅
• ::1 ✅
• Any remote host ❌ Rejected

The LocalOnlyPolicy class enforces this at the backend level.

---

License

MIT