antiai 0.1.0

Advanced image protection system against AI training and unauthorized use

AntiAI - Advanced Image Protection System

AntiAI is a comprehensive Python library for protecting images against unauthorized AI training, style mimicry, and content theft. It implements state-of-the-art adversarial perturbations, invisible watermarking, and robust metadata management.

🎯 Key Features

• Adversarial Protection: Imperceptible perturbations that confuse AI models
• Invisible Watermarking: Robust watermarks surviving compression and edits
• Metadata Management: Complete authorship and copyright tracking
• File Integrity: Cryptographic signatures and tampering detection
• Custom Format: .antiAI format designed specifically for protection
• Easy Integration: Simple API for encoding, decoding, and verification

🚀 Quick Start

Installation

pip install antiai

Basic Usage

from antiai import AntiAIEncoder, AntiAIDecoder

# Protect an image
encoder = AntiAIEncoder()
encoder.encode(
    input_image="artwork.jpg",
    output_path="artwork.antiAI",
    author="Your Name",
    protection_level=7  # 0-10
)

# Decode protected image
decoder = AntiAIDecoder()
image, metadata = decoder.decode("artwork.antiAI")

print(f"Author: {metadata.author.name}")
print(f"Protected: {metadata.protection.adversarial}")

📚 Documentation

Protection Mechanisms

1. Adversarial Perturbation

Adds imperceptible noise that causes AI models to fail at:

• Feature extraction
• Style transfer
• Content recognition

from antiai.protection import AdversarialProtection

protector = AdversarialProtection(strength=8)
protected_image, metadata = protector.protect(original_image)

print(f"PSNR: {metadata['quality']['psnr_db']:.2f} dB")  # >40 = imperceptible

2. Invisible Watermarking

Embeds identifying information using DWT-DCT:

• Survives JPEG compression
• Resistant to resizing
• Extractable for proof of ownership

from antiai.protection import InvisibleWatermark

watermarker = InvisibleWatermark(strength=0.1)
marked_image, metadata = watermarker.embed(image, "unique_id_12345")

3. Metadata & Copyright

Complete authorship tracking with:

• Creator information
• Copyright statements
• Usage terms
• AI training restrictions

from antiai.protection.metadata import MetadataBuilder

metadata = (
    MetadataBuilder()
    .set_author("Artist Name", email="[e-mail]")
    .set_copyright("© 2025 Artist Name", license="All Rights Reserved")
    .set_ai_restrictions(do_not_train=True, do_not_scrape=True)
    .build()
)

File Format

The .antiAI format structure:

[HEADER - 512 bytes]
  - Magic bytes: "ANTIAI\x00\x01"
  - Version, dimensions, protection level
  
[METADATA CHUNK]
  - Author, copyright, creation date
  - Protection details
  
[WATERMARK CHUNK]
  - Invisible watermark data
  
[IMAGE DATA CHUNK]
  - PNG-compressed protected image
  
[SIGNATURE - 256 bytes]
  - SHA-256 integrity signature

🔍 Advanced Features

Batch Processing

encoder = AntiAIEncoder()
results = encoder.encode_batch(
    input_images=["img1.jpg", "img2.jpg", "img3.jpg"],
    output_dir="protected/",
    author="Batch User",
    protection_level=6
)

Verification

from antiai import ProtectionVerifier

verifier = ProtectionVerifier()
result = verifier.verify("artwork.antiAI")

if result['authentic']:
    print("✓ Image is authentic and protected")
else:
    print(f"✗ Issues: {result['issues']}")

Image Comparison

verifier = ProtectionVerifier()
comparison = verifier.compare_images("original.jpg", "copy.jpg")

print(f"Similarity: {comparison['similarity_percent']:.1f}%")
print(f"Likely same image: {comparison['likely_same_image']}")

🛠️ Tools

The examples/ folder includes ready-to-use tools:

Tool	Description
viewer.py	Minimalist GUI viewer for .antiAI files (Tkinter)
protect.py	CLI tool for protecting images
demo.py	Interactive demonstration of all features

# View a protected image
python examples/viewer.py

# Protect an image from command line
python examples/protect.py input.jpg -o output.antiAI --level 7

# Run interactive demo
python examples/demo.py

🛡️ Protection Levels

Level	Strength	Use Case	Quality Impact
0-2	Minimal	Basic protection	Imperceptible
3-5	Moderate	General use	Very minor
6-8	Strong	Professional work	Minor
9-10	Maximum	High-value art	Noticeable

📊 Performance

Typical encoding performance on a modern CPU:

• Small images (512x512): ~1-2 seconds
• Medium images (1920x1080): ~3-5 seconds
• Large images (4K): ~10-15 seconds

GPU acceleration (CUDA) can provide 2-4x speedup.

🔬 Technical Details

Adversarial Algorithm

• Method: Projected Gradient Descent (PGD)
• Loss Function: Feature maximization
• Constraint: L-infinity bounded perturbations
• Quality: PSNR > 40 dB typical

Watermark Algorithm

• Transform: Discrete Wavelet Transform (DWT)
• Embedding: DCT coefficient modification
• Robustness: Mid-frequency embedding
• Capacity: ~0.1 bits per pixel

🤝 Contributing

Contributions are welcome! Please see CONTRIBUTING.md for details.

Development Setup

git clone https://github.com/Ross-cripto/antiai.git
cd antiai
pip install -e ".[dev]"
pre-commit install

Running Tests

pytest test/ -v --cov=antiai

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• Inspired by Glaze and Nightshade
• Built with PyTorch, NumPy, and PIL
• Special thanks to the computer vision and adversarial ML research communities

📧 Contact

• Author: Rosniel Miguel Allesta Fundora
• Email:[e-mail]
• GitHub: @Ross-cripto

🔗 Links

• Issue Tracker

---

Made with ❤️ to protect artists and creators from unauthorized AI use.