FairQuanti: Enhancing Fairness in Deep Neural Network Quantization via Neuron Role Contribution

This repository contains the implementation of the paper "FairQuanti: Enhancing Fairness in Deep Neural Network Quantization via Neuron Role Contribution". The project focuses on improving fairness in deep neural network quantization by considering the role and contribution of individual neurons.

---

📁 Project Structure

• train.py: Script for training the neural network models on fairness-sensitive datasets.
• main.py: Script for implementing the FairQuanti method and evaluating its performance.
• evaluate.py: Utilities for model evaluation and metrics computation.
• models.py: Definitions of neural network architectures (e.g., Net, Net_CENSUS, NetPlus_).
• utils/: Helper functions for dataset transformation and preprocessing.
• data/: Directory for storing datasets (e.g., COMPAS, Census, Credit).
• results/: Directory for saving trained models and outputs.

---

🛠 Requirements

Make sure you have Python 3.8+ installed. Install dependencies with:

pip install -r requirements.txt

Datasets

The code supports three datasets:

• COMPAS: Recidivism prediction dataset.
• Census: Adult income prediction dataset.
• Credit: German credit risk dataset.

Place the dataset files in the data/ directory with the following structure:

data/
├── COMPAS/compas_recidive_two_years_sanitize_age_category_jail_time_decile_score.csv
├── Census/adult
└── Credit/german_credit

Usage

1. Training the Model (train.py)

Use train.py to train a neural network on a specified dataset. Below are example commands:

• Train on COMPAS dataset:

  python train.py --dataset compas --batch-size 128 --bit 8 --input-path ./data/COMPAS/compas_recidive_two_years_sanitize_age_category_jail_time_decile_score.csv --save-dir ./results

• Train on Census dataset:

  python train.py --dataset census --batch-size 128 --bit 8 --input-path ./data/Census/adult --save-dir ./results

• Train on Credit dataset:

  python train.py --dataset credit --batch-size 128 --bit 8 --input-path ./data/Credit/german_credit --save-dir ./results

2. Implementing FairQuanti (main.py)

Use main.py to implement the FairQuanti method and evaluate its performance. Below are example commands:

• Run FairQuanti on COMPAS dataset:

  python main.py --dataset compas --batch-size 128 --bits 8 --input-path ./data/COMPAS/compas_recidive_two_years_sanitize_age_category_jail_time_decile_score.csv --save-dir ./results

• Run FairQuanti on Census dataset:

  python main.py --dataset census --batch-size 128 --bits 8 --input-path ./data/Census/adult --save-dir ./results

• Run FairQuanti on Credit dataset:

  python main.py --dataset credit --batch-size 128 --bits 8 --input-path ./data/Credit/german_credit --save-dir ./results

Results

The results are saved in the results/ directory. The results include the trained model, the quantized model, and the evaluation metrics.

Citation

If you find this work useful, please cite it as follows:

@article{fairquanti,
  title={FairQuanti: Enhancing Fairness in Deep Neural Network Quantization via Neuron Role Contribution},
  author={Zhiqi Cao and Jingjing Li and Jingwei Li and Jingren Zhou},
  year={2025},
}