DPFairDeepLearning

This repository contains code to implement differentially private fair deep learning models for tabular data. The code allows training models with different architectures, datasets, and privacy guarantees.

Models

The following model architectures are implemented:

• DemParModel - Implements statistical parity via an adversarial network
• EqualOddModel - Extends DemParModel for equalized odds via additional input to adversary

The models consist of an autoencoder, classifier, and adversary. Differential privacy can be applied to the encoder+classifier or adversary components.

Datasets

The following datasets are used:

• Adult
• German
• CelebA

Preprocessing code is provided for each dataset to extract features, labels, and sensitive attributes.

Training

The main training code is in run_training.py. This handles model instantiation, data loading, and trainer initialization.

The Trainer class in trainer.py handles the full training loop including:

• Making models differentially private
• Training autoencoder + classifier
• Adversarial training loop
• Logging metrics
• Evaluating fairness metrics
• Checkpointing based on accuracy and fairness

The training_script.py shows how the hyperparameters and model configurations are defined. It generates the full set of experiments and executes them, leveraging parallel threads.

Utilities

utils.py contains utility functions like:

• Converting data to torch Tensors
• Generating config file strings for ClearML
• Logging metrics to ClearML
• Filtering hyperparameter combinations

Notebooks

The data_management folder contains Jupyter notebooks for saving/loading experiments from ClearML.

Requirements

The code requires Python 3 and the following libraries:

• PyTorch
• ClearML
• Scikit-learn
• Numpy
• Pandas

Install requirements with:

pip install -r requirements.txt

Running Experiments

To run an example experiment:

python run_training.py --dataset Adult --model DP --epochs 10 

This will train a DemPar model on the Adult dataset for 10 epochs.

See training_script.py for the full list of possible hyperparameters.

Instructions

To run an experiment:

1. Call run_training.py for each desired configuration or generate and execute list of hyperparameter configurations in training_script.py
2. Monitor experiments in ClearML Web UI
3. Use notebooks to collect results

So in summary, the key files are:

• run_training.py - Main training loop
• trainer.py - Full training implementation
• training_script.py - Hyperparameter generation and execution main loop
• utils.py - Helper functions
• data_management/ - Notebooks for ClearML

The other files provide dataset processing, model architectures, metrics, etc.