Fairness of Compression Techniques in CNNs

This repository accompanies our mid-project report and final poster on investigating the interplay between model compression and algorithmic fairness in convolutional neural networks. We focus on quantization-based compression and two post-hoc bias mitigation strategies:

• Calibrated Equalized Odds (CalEqOd)
• FairALM (Augmented Lagrangian Method)

📄 Mid-Project Report

A detailed write-up of our methodology, datasets, and preliminary approach is available in the PDF:
Fairness_and_Compression.pdf

Final Poster

Our poster summarizes key findings, including compression ratios and fairness trade-offs:
View Poster (PDF)

---

Requirements & Setup

• Python: ≥3.8
• Install dependencies

  pip install -r requirements.txt

---

Repository Structure

.
├── data/                   # CelebA subsets and preprocessing scripts
├── src/                    # Training, quantization & evaluation code
│   ├── train.py            # Model training with/without fairness constraints
│   ├── quantize.py         # Post-training quantization routines
│   └── eval.py             # Compute accuracy, EO, FPR disparities
├── trials/                 # Saved checkpoints & log files
├── Fairness_and_Compression.pdf  # Mid-project report
├── fairness_of_model_poster_aryan.pdf  # Final poster
├── requirements.txt
└── README.md

---

Usage

1. Train & Evaluate (baseline CNN)

   python src/train.py \
     --dataset celeba \
     --epochs 20 \
     --save-dir trials/baseline/
   python src/eval.py \
     --checkpoint trials/baseline/model.pt

2. Quantize & Measure Fairness

   python src/quantize.py \
     --checkpoint trials/baseline/model.pt \
     --output-dir trials/quantized/
   
   python src/eval.py \
     --checkpoint trials/quantized/model_int8.pt

3. Apply Bias Mitigation

   • Calibrated Equalized Odds

     python src/train.py \
       --dataset celeba \
       --epochs 10 \
       --fairness caleqod \
       --save-dir trials/caleqod/

   • FairALM

     python src/train.py \
       --dataset celeba \
       --epochs 10 \
       --fairness fairalm \
       --save-dir trials/fairalm/

---

Key Results

• Bias mitigation: Implemented post-training Calibrated Equalized Odds and FairALM, reducing demographic disparity by 9.5 % (EO gap 0.21→0.19) while maintaining accuracy (82 %→79 %).
• Compression: Achieved 4× model size reduction via post-training quantization with minimal fairness degradation.

---

Future Prospects:

• Explore joint optimization of compression and fairness constraints during training.
• Evaluate on additional protected attributes and datasets.
• Integrate more advanced compression schemes (e.g., pruning + quantization).

---