veridex 0.1.2

A modular, probabilistic, and research-grounded AI content detection library.

Veridex

A modular, probabilistic, and research-grounded AI content detection library.

Veridex is a production-ready library for detecting AI-generated content across multiple modalities: text, image, and audio. Unlike binary classifiers, Veridex provides probabilistic detection with confidence estimates and interpretable signals.

✨ Features

• 🎯 Multi-Modal Detection: Text, Image, and Audio deepfake detection
• 📊 Probabilistic Outputs: Returns probabilities and confidence scores, not just binary labels
• 🔍 Interpretable Signals: Exposes individual detection features for transparency
• 🧩 Modular Architecture: Easy to extend with new detection methods
• 🚀 Production-Ready: Robust error handling, graceful degradation
• 📖 Research-Grounded: Based on state-of-the-art papers and benchmarks

🚀 Quick Start

Installation

# Install core library
pip install veridex

# Install with specific modality support
pip install veridex[text]      # Text detection
pip install veridex[audio]     # Audio detection
pip install veridex[image]     # Image detection

# Install everything
pip install veridex[text,image,audio]

# Development installation
pip install -e ".[dev]"

Usage Examples

Text Detection

from veridex.text import PerplexitySignal, BinocularsSignal

# Quick detection with perplexity
detector = PerplexitySignal()
result = detector.run("This text seems suspiciously perfect...")

print(f"AI Probability: {result.score:.2f}")
print(f"Confidence: {result.confidence:.2f}")
print(f"Perplexity: {result.metadata['mean_perplexity']:.2f}")

Audio Detection

from veridex.audio import SpectralSignal

# Lightweight frequency analysis
detector = SpectralSignal()
result = detector.run("audio_sample.wav")

print(f"AI Probability: {result.score:.2f}")
print(f"Spectral Features: {result.metadata}")

Image Detection

from veridex.image import FrequencyDomainSignal

# Analyze spectral anomalies
detector = FrequencyDomainSignal()
result = detector.run("suspicious_image.png")

print(f"AI Probability: {result.score:.2f}")

📦 Available Detectors

Text Detectors

Detector	Method	Speed	Accuracy	GPU Required
ZlibEntropySignal	Compression-based	Fast	Low	No
PerplexitySignal	Statistical (LLM-based)	Medium	Medium	Optional
BinocularsSignal	Contrastive Perplexity	Medium	High	Optional

Audio Detectors

Detector	Method	Speed	Accuracy	GPU Required
SpectralSignal	Frequency Domain	Fast	Medium	No
AASISTSignal	Spectro-Temporal	Medium	High	No
Wav2VecSignal	Foundation Model	Slow	Very High	Recommended

Image Detectors

Detector	Method	Speed	Accuracy	GPU Required
FrequencyDomainSignal	Spectral Analysis	Fast	Medium	No
DIRESignal	Diffusion Reconstruction	Slow	High	Yes

🧪 Testing

# Create virtual environment
python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install with development dependencies
pip install -e ".[dev,text,audio,image]"

# Run all tests
pytest tests/ -v

# Run specific module tests
pytest tests/audio/ -v

# With coverage
pytest tests/ --cov=veridex --cov-report=html

See TESTING.md for detailed testing guide.

📚 Documentation

• Core Idea & Plan - Design philosophy and architecture
• Research Document - Comprehensive technical analysis
• Testing Guide - How to test all metrics
• Module READMEs:
  • Audio Detection
  • Examples in examples/ directory

🏗️ Architecture

Veridex follows a signal-based architecture:

Input → Signal Extractors → Normalization → Fusion → Output
                ↓
    (Independent, Inspectable Signals)

Each detector:

• Inherits from BaseSignal
• Returns standardized DetectionResult
• Operates independently
• Declares its limitations explicitly

🤝 Contributing

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

Development Setup

# Clone repository
git clone https://github.com/ADITYAMAHAKALI/veridex.git
cd veridex

# Create virtual environment
python3 -m venv venv
source venv/bin/activate

# Install in editable mode with all dependencies
pip install -e ".[dev,text,image,audio]"

# Run tests
pytest tests/

# Format code
black veridex/ tests/
flake8 veridex/

📄 License

Apache License 2.0 - See LICENSE for details.

🔬 Research & Citations

Veridex is based on cutting-edge research in AI-generated content detection. Key methods include:

• Binoculars: Spotting LLMs With Binoculars (arXiv:2401.12070)
• AASIST: Audio Anti-Spoofing Integrated Spectro-Temporal Graph Attention
• DIRE: Diffusion Reconstruction Error for deepfake images
• Wav2Vec 2.0: Self-supervised foundation models for audio

See docs/research.md for full references.

⚠️ Limitations

Veridex is a probabilistic detection tool, not a definitive proof system:

• ❌ Not suitable as sole evidence for legal/forensic purposes
• ❌ Cannot detect all AI-generated content with 100% accuracy
• ❌ Vulnerable to adversarial attacks and post-processing
• ⚠️ Requires regular updates as generative models improve

Always use multiple signals and human judgment for critical decisions.

🗺️ Roadmap

• Text detection (Perplexity, Binoculars)
• Image detection (Frequency, DIRE)
• Audio detection (Spectral, AASIST, Wav2Vec)
• Video detection (rPPG, I3D)
• C2PA provenance integration
• Ensemble fusion models
• Real-time streaming detection
• Model calibration on benchmarks

📧 Contact

For questions, issues, or contributions:

• Email: [e-mail]
• Open an issue on GitHub
• See CONTRIBUTING.md

---

Built with ❤️ for transparency in the age of generative AI