⚖️ FairFace: Debiasing Facial Detection using DB-VAE

📌 Project Overview

FairFace is a Deep Learning project aimed at identifying and mitigating algorithmic bias in facial detection systems. Standard CNNs often perform poorly on underrepresented groups (e.g., darker skin tones or specific gender/age groups) due to dataset imbalance.

This project implements a Debiasing Variational Autoencoder (DB-VAE) to learn the latent structure of facial features and automatically re-weight training examples during learning, ensuring fair classification accuracy across all demographics.

⚙️ The Problem & Solution

• The Bias: Models trained on skewed datasets (like CelebA) tend to learn features correlated with the majority class (e.g., Light Skin / Female), leading to high error rates for minorities.
• The Fix (DB-VAE):
  1. Latent Learning: A VAE learns a low-dimensional representation of faces (Latent Space).
  2. Automated Re-sampling: The model identifies images in "sparse" regions of the latent space (underrepresented faces) and increases their sampling probability during training.
  3. Result: The classifier sees "hard" or "rare" examples more often, reducing bias without needing manual labeling of attributes.

🛠️ Technical Implementation

• Architecture: Convolutional VAE (Encoder-Decoder) with a Classification Head.
• Loss Function:
  • VAE Loss: Reconstruction Loss + KL Divergence (for latent learning).
  • Classification Loss: Cross-Entropy (for face detection).
• Debiasing Algorithm: Adaptive re-sampling based on the estimated latent distribution $Q(z|x)$.

🚀 How to Run

1. Clone the repository:

   git clone [https://github.com/mariamashraf731/FairFace-DBVAE.git](https://github.com/mariamashraf731/FairFace-DBVAE.git)

2. Install Dependencies:

   pip install -r requirements.txt

3. Run the Analysis:

   jupyter notebook notebooks/Debiasing_Facial_Detection.ipynb

👨‍💻 Tech Stack

• Language: Python
• Deep Learning: TensorFlow / Keras
• Concepts: Unsupervised Learning, VAEs, Latent Space Analysis, Algorithmic Fairness.