An interactive dashboard for evaluating Responsible AI properties of machine learning models.
The Ethical Dashboard allows researchers and developers to test machine learning models across four critical ethical dimensions:
- Fairness
- Robustness
- Privacy Preservation
- Explainability
The tool provides an end-to-end interface to upload datasets, train or upload models, and analyze the ethical implications of model predictions.
A live demo of the dashboard is available here:
https://ethical-dashboard.streamlit.app/
You can also use how_to_use_the_app.md as an interactive guide for "how-to-use".
- Upload custom datasets (CSV format)
- Use built-in benchmark datasets
- Visualize dataset distributions and statistics
- Edit datasets directly inside the dashboard
Users can:
- Select predefined models
- Upload their own trained models
Supported model formats include:
.pkl.h5.hdf5.safetensors
Supported frameworks:
- scikit-learn
- TensorFlow / Keras
- PyTorch
The dashboard automatically inspects models and displays:
- model type
- hyperparameters
- architecture summary (when available)
After training or loading a model, the dashboard evaluates it using four Responsible AI modules.
Evaluates prediction disparities across sensitive groups.
Metrics include:
- Accuracy per group
- Precision
- False Positive Rate (FPR)
- False Negative Rate (FNR)
- Selection rate
- Group population statistics
This helps detect potential bias in model predictions.
Measures how model performance changes under input perturbations.
The dashboard evaluates robustness by applying:
- Gaussian noise to inputs
- Perturbation-based stress testing
Outputs include:
- accuracy degradation curves
- robustness visualizations
This helps detect models that are overly sensitive to small changes in input data.
Evaluates privacy risks using Membership Inference Attacks (MIA).
The module estimates whether an attacker can infer whether a sample was used in the training dataset.
Metrics include:
- attack accuracy
- attack precision
- attack recall
- attack F1-score
- attack confusion matrix
This helps estimate training data leakage risk.
Uses SHAP (SHapley Additive exPlanations) to explain model predictions.
The module provides:
- global feature importance
- local explanations for individual predictions
- explanation stability metrics
- explanation sparsity metrics
This helps interpret which features influence model predictions.
git clone https://github.com/DevBhuyan/Ethical-Dashboard.git
cd Ethical-Dashboard
python -m pip install -r requirements.txt
Or if you prefer uv
uv pip install -r requirements.txt
streamlit run app.py
Then open a browser at http://localhost:8501
You must see the landing page as such:
Happy evaluating!
For more information on how to use the app, see how_to_use_the_app.md