TrustLens: Explainable & Trust-Aware AI 🔍

TrustLens is an explainable and trust-aware AI prototype that classifies images while revealing how confident the model really is. Built for transparency and reliability in AI decision-making.

🚀 Features

• Image Classification: CNN trained on CIFAR-10 with 10 classes
• Uncertainty Quantification: Entropy-based confidence scoring
• Anomaly Detection: Multiple methods including ODIN and Mahalanobis distance
• Explainable AI: Grad-CAM visualizations showing decision reasoning
• Trust Calibration: Temperature scaling for confidence calibration
• Interactive Interface: Streamlit-based web application
• Real-time Analysis: Upload images and get instant predictions with explanations

🛠️ Installation

Prerequisites

• Python 3.8+
• PyTorch
• CUDA (optional, for GPU acceleration)

Setup

1. Clone the repository:

git clone https://github.com/Leptons1618/TrustNet.git
cd TrustNet

2. Create a virtual environment:

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies:

pip install -r requirements.txt

4. Download CIFAR-10 data (will be downloaded automatically on first run)

🎯 Usage

Running the Application

streamlit run app.py

The application will open in your default web browser at http://localhost:8501.

Features Overview

1. Image Upload & Classification

• Upload any image or select from sample images
• Get predicted class with confidence score
• View preprocessing steps

2. Trust Analysis

• Entropy Score: Measures prediction uncertainty
• ODIN Score: Out-of-distribution detection
• Mahalanobis Distance: Anomaly detection based on feature space
• Calibrated Confidence: Temperature-scaled probability

3. Explainability

• Grad-CAM Heatmaps: Visual explanation of model decisions
• Feature Importance: Which parts of the image matter most
• Decision Reasoning: Step-by-step explanation

4. Trust Dashboard

• Real-time trust metrics
• Historical analysis
• Confidence calibration plots

🏗️ Architecture

TrustNet/
├── src/
│   ├── models/
│   │   ├── __init__.py
│   │   ├── cnn.py              # CNN architecture
│   │   └── trust_model.py      # Trust-aware model wrapper
│   ├── trust_methods/
│   │   ├── __init__.py
│   │   ├── entropy.py          # Entropy-based uncertainty
│   │   ├── odin.py            # ODIN method
│   │   ├── mahalanobis.py     # Mahalanobis distance
│   │   └── temperature.py     # Temperature scaling
│   ├── utils/
│   │   ├── __init__.py
│   │   ├── data_loader.py     # Data preprocessing
│   │   ├── gradcam.py         # Grad-CAM implementation
│   │   └── visualization.py   # Plotting utilities
│   └── app.py                 # Main Streamlit application
├── assets/                    # Static files and sample images
├── models/                    # Trained model files
├── tests/                     # Unit tests
├── cookbook/                  # Jupyter notebooks for development
├── requirements.txt
├── README.md
└── .gitignore

🔬 Trust Methods

1. Entropy-Based Uncertainty

Measures prediction uncertainty using Shannon entropy:

H(p) = -∑ p_i * log(p_i)

• Low entropy = High confidence
• High entropy = High uncertainty

2. ODIN (Out-of-Distribution Detection)

Uses input preprocessing and temperature scaling to detect anomalous inputs:

• Applies small perturbations to inputs
• Uses temperature scaling on logits
• Effective for detecting domain shift

3. Mahalanobis Distance

Measures distance from class centroids in feature space:

• Computes distance to nearest class centroid
• Uses feature representations from penultimate layer
• Effective for detecting novel classes

4. Temperature Scaling

Calibrates model confidence to match actual accuracy:

• Post-processing technique
• Learns optimal temperature parameter
• Improves reliability of confidence scores

📊 Model Performance

CIFAR-10 Classification

• Clean Accuracy: ~85%
• Domain-Shifted Accuracy: ~70%
• Calibration Error: <5% (after temperature scaling)

Trust Methods Performance

• Entropy AUC: 0.85+ for OOD detection
• ODIN AUC: 0.90+ for OOD detection
• Mahalanobis AUC: 0.88+ for OOD detection

🎨 Sample Use Cases

1. Quality Inspection: Detect defective products with confidence scores
2. Medical Triage: Classify medical images with uncertainty quantification
3. Security Screening: Identify suspicious items with explainable decisions
4. Autonomous Systems: Make safety-critical decisions with trust metrics

🧪 Development

Running Tests

pytest tests/

Training New Models

python src/train.py --config config/train_config.yaml

Jupyter Notebooks

Explore the cookbook/ directory for development notebooks and experiments.

📈 Roadmap

• Support for additional datasets (ImageNet, custom datasets)
• Ensemble methods for improved trust
• Bayesian neural networks
• Active learning integration
• Mobile app deployment
• REST API for integration

🤝 Contributing

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Commit your changes (git commit -m 'Add amazing feature')
4. Push to branch (git push origin feature/amazing-feature)
5. Open a Pull Request

📄 License

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

🙏 Acknowledgments

• CIFAR-10 dataset creators
• PyTorch team for the deep learning framework
• Streamlit team for the web app framework
• Research papers on uncertainty quantification and explainable AI

📧 Contact

For questions or support, please open an issue on GitHub or contact [anishgiri163@gmail.com].

---

Made with ❤️ for transparent and trustworthy AI