Hansa

Deep learning-based binary classifier for detecting AI-generated images. Achieves 88%+ accuracy on 32x32 RGB images using a lightweight CNN architecture.

Why Hansa?

Hansa (Sanskrit: हंस) refers to the sacred swan in Hindu mythology, renowned for its mythical ability to separate milk from water when mixed together. This symbolizes the power of discernment and the ability to distinguish truth from falsehood—precisely what this model does by separating real images from AI-generated ones.

Performance

• Test Accuracy: 88.34%
• Precision: ~87.44%
• Recall: ~89.56%
• F1-Score: ~88.48%
• AUC: 0.95+
• Inference Time: <10ms per image

Architecture

Input (32x32x3)
    ↓
Data Augmentation (flip, rotation, zoom)
    ↓
3 Conv Blocks (32→64→128 filters)
    ↓
Global Average Pooling
    ↓
Dense (128) + Dropout (0.5)
    ↓
Output (sigmoid)

Key Features:

• Batch Normalization after each conv layer
• L2 Regularization (0.001)
• Dropout (0.2-0.5)
• Adam optimizer (LR: 1e-5)

Requirements

tensorflow>=2.10.0
numpy>=1.21.0
matplotlib>=3.5.0
seaborn>=0.11.0
scikit-learn>=1.0.0

Installation

# Clone repository
git clone <repository-url>
cd hansa

# Create virtual environment
python -m venv hansa_env
source hansa_env/bin/activate  # On Windows: hansa_env\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Dataset Structure

data/
├── train/
│   ├── REAL/     # 50,000 real images
│   └── FAKE/     # 50,000 AI-generated images
└── test/
    ├── REAL/     # 10,000 real images
    └── FAKE/     # 10,000 AI-generated images

Image Specifications:

• Format: JPG/PNG
• Size: 32x32 pixels
• Channels: RGB (3)
• Classes: Binary (REAL/FAKE)

Training

jupyter notebook hansa.ipynb

Training Configuration:

• Epochs: 20 (with early stopping)
• Batch Size: 32
• Validation Split: 20%
• Callbacks: Early stopping, model checkpoint, LR reduction

Outputs:

• hansa_best.keras - Best model checkpoint
• hansa.keras - Final model
• training_config.json - Training parameters and metrics
• training_history.png - Loss/accuracy curves
• confusion_matrix.png - Confusion matrix + ROC curve
• sample_predictions.png - Prediction examples

Inference

Single Image

python predict.py image.jpg

Output:

✅ REAL IMAGE DETECTED
   Confidence: 87.34%

Verbose Mode

python predict.py image.jpg --verbose

Output:

✅ REAL IMAGE DETECTED
   Confidence: 87.34%
   Real Score: 87.34%
   Fake Score: 12.66%
   Threshold: 0.5
   Inference Time: 8.42ms
   Image: image.jpg

JSON Output

python predict.py image.jpg --json

Output:

{
  "image_path": "image.jpg",
  "predicted_class": "REAL",
  "confidence": 0.8734,
  "real_score": 0.8734,
  "fake_score": 0.1266,
  "threshold": 0.5,
  "inference_time_ms": 8.42
}

Batch Processing

python predict.py ./images --batch

Output:

==================================================
BATCH PREDICTION SUMMARY
==================================================
Total Images: 100
Successfully Processed: 100
Errors: 0

✅ Real Images: 52 (52.0%)
⚠️  Fake Images: 48 (48.0%)

Average Confidence: 83.45%
==================================================

Custom Threshold

python predict.py image.jpg --threshold 0.7

All Options

python predict.py <path> [options]

Options:
  --json              Output in JSON format
  --batch             Process directory of images
  --threshold FLOAT   Classification threshold (default: 0.5)
  --verbose           Show detailed output
  --quiet             Minimal output
  --model PATH        Path to model file (default: hansa.keras)

Model Files

• hansa.keras (3.2 MB) - Trained model
• training_config.json - Model metadata and metrics

Evaluation Metrics

The model provides:

• Accuracy: Overall correctness
• Precision: Of detected fakes, how many are actually fake
• Recall: Of all fakes, how many were detected
• F1-Score: Harmonic mean of precision and recall
• AUC: Area under ROC curve

Project Structure

hansa/
├── Hansa.ipynb                   # Training notebook
├── predict.py                    # Inference script
├── hansa.keras                   # Trained model
├── hansa_best.keras              # Best trained model
├── training_config.json          # Training metadata
├── requirements.txt              # Dependencies
├── README.md                     # Documentation
└── data/                         # Dataset (not included)
    ├── train/
    └── test/

Hardware Requirements

Training:

• RAM: 8GB minimum
• GPU: Optional (Apple M1/M2, NVIDIA CUDA)
• Storage: 5GB for dataset

Inference:

• RAM: 2GB minimum
• CPU: Any modern processor
• GPU: Not required

Limitations

• Trained on 32x32 images only
• Binary classification (no multi-class)
• Performance degrades on high-resolution images
• Limited to RGB images
• Dataset-specific (may not generalize to all AI generators)

Future Improvements

• Support for higher resolutions (128x128, 256x256)
• Multi-class detection (identify AI model used)
• Transfer learning from larger models
• Ensemble methods
• Real-time video processing
• Web API deployment

License

MIT License - See LICENSE file for details

Citation

If you use this model in your research, please cite:

@software{hansa2025,
  title={Hansa},
  author={tyrobro},
  year={2025},
  url={https://github.com/tyrobro/Hansa}
}

Contact

For issues, questions, or contributions:

• GitHub Issues: Create Issue

---

Last Updated: November 2025
Model Version: 1.0
Framework: TensorFlow 2.x