This repository contains the official codebase for our study on bias mitigation in soft facial attribute recognition.

We present a fully automated framework for the targeted generation of synthetic facial images under controlled demographic conditions.
Our method integrates both prompt-based (diffusion) and latent-space (GAN) generative models to synthesize demographically balanced data, which is automatically annotated with facial attributes and demographic labels.
This enables a rigorous, fine-grained fairness evaluation across gender, age, and ethnicity.

Experiments conducted on the MAAD-Face benchmark show that our method significantly reduces disparities between demographic groups—without sacrificing overall classification performance.
The complete dataset, code, and evaluation tools are released to support reproducible and fairness-aware research.

📄 Read the paper (IEEE)

synthetic-fairness-faces is a modular and fully automated framework for generating and annotating synthetic facial datasets aimed at improving fairness in soft-biometric attribute recognition.

We designed the system to integrate Stable Diffusion, StyleGAN3, Pix2Pix Diffuser, MagFace, FaceXFormer, FairFace, and Slim-CNN into a cohesive pipeline for image generation, attribute annotation, classifier training, and demographic fairness evaluation.

• 🧬 Dual image generation: prompt-based (Stable Diffusion) and latent-based (GAN)
• 🎯 Attribute control: generate faces by gender, age, ethnicity, and hair color
• 🧹 Quality filtering: discard low-quality or unrealistic faces using MagFace embeddings
• 🏷️ Automated annotation: soft-biometric attributes (FaceXFormer) and demographics (FairFace)
• 🧪 Fairness evaluation: compute FDR, Demographic Parity, and Equalized Odds across subgroups
• 📊 Model retraining: evaluate Slim-CNN on real + synthetic data (CelebA / MAAD-Face)

• Stable Diffusion (Realistic Vision V5.0 noVAE) for text-to-image generation
• StyleGAN3 + Pix2Pix Diffuser for GAN-based latent manipulation and refinement
• MagFace for identity embeddings and quality-aware filtering
• FaceXFormer and FairFace for multi-attribute and demographic labeling
• Slim-CNN: efficient model for soft attribute classification
• CelebA and MAAD-Face as benchmark datasets
• Python, PyTorch, Hugging Face, OpenCV, dlib, NumPy

The synthetic-fairness-faces framework supports the full bias mitigation workflow with a sequence of automated steps:

1. Prompt checking and validation

   • Ensures the prompt.txt file is present and non-empty

2. Synthetic image generation

   • Default: Stable Diffusion (generate_images_with_selected_model.py)
   • Alternative: GAN-based methods (manual switch required)

3. Image resizing and quality filtering

   • Resize to 512x512, 178x218, and 112x112
   • Filter low-quality faces using MagFace magnitude

4. Feature extraction and evaluation

   • Extract MagFace embeddings
   • Discard samples with magnitude < 20

5. Attribute and demographic labeling

   • Two options:
     • Facer + FairFace: label with Facer, validate with FairFace
     • FaceXFormer: single-step labeling via transformer model

6. Dataset partitioning and model training

   • Partition into train/val/test (80/10/10)
   • Train Slim-CNN on CelebA + synthetic data

7. Evaluation and fairness analysis

   • Evaluate Slim-CNN on MAAD-Face benchmark
   • Analyze performance across gender, age, and ethnicity

You can execute the entire process automatically by running pipeline_automatic.py.
The script performs all steps in sequence, including image generation, preprocessing, labeling, training, and fairness evaluation.

A detailed file tree is available here

Install all required dependencies using Conda:

conda env create -f environment.yml
conda activate env

Model and dataset requirements are automatically checked at runtime by pipeline_automatic.py.
If any file is missing, the script will print the direct download links.

• vae.pt for image generation (Stable Diffusion VAE)
  Download → place in: models/image_generation/

• FaceXFormer
  Download → place in: models/facexformer/

• FairFace models → place all 3 files in: models/FairFace/

  • mmod_human_face_detector.dat
  • shape_predictor_5_face_landmarks.dat
  • res34_fair_align_multi_4_20190809.pt

• MagFace model
  Download → place in: models/magface/

• CelebA dataset
  Download
  Extract to: img/CelebA/img_align_celeba/

• VGGFace2 (test set only)
  Download on Kaggle
  Extract only the val/ folder to: img/vggface2_test/test/

• MAAD_Face.csv
  Download
  Place in: img/vggface2_test/

To execute the full workflow:

python pipeline_automatic.py

You’ll be prompted to select the labeling method:

• 1 → Facer + FairFace (separate labeling + demographic validation)
• 2 → FaceXFormer (joint attribute and demographic labeling)

By default, image generation is performed using Stable Diffusion (SD).

🌀 To use GAN-based generation instead, comment the SD script inside pipeline_automatic.py and enable the appropriate one in:

image_generation/src/GAN/

All downstream modules (resizing, filtering, annotation, training, and evaluation) remain compatible regardless of the generator used.

• 🔧 Training logs and checkpoints:

  pytorch-slim-cnn/checkpoints*
  pytorch-slim-cnn/plots/
  

• 📓 Evaluation Notebook (opens in Jupyter):

  pytorch-slim-cnn/evaluate_both.ipynb
  

• 📄 Fairness metrics and predictions:

  fairness/*.csv
  

• Mattia Maucioni | 📧 mattiamaucioni [at] icloud.com | Data Science Field
• Marco Di Maio | 📧 marcodimaio2102 [at] gmail.com | Data Science Field

L. Cascone, M. Di Maio, V. Loia, M. Maucioni, M. Nappi, and C. Pero,
“Synthetic Data for Fairness: Bias Mitigation in Facial Attribute Recognition,”
IEEE Open Journal of the Computer Society, vol. 6, pp. 1703–1714, 2025.
doi: 10.1109/OJCS.2025.3622694

@ARTICLE{11206468,
  author={Cascone, Lucia and Maio, Marco Di and Loia, Vincenzo and Maucioni, Mattia and Nappi, Michele and Pero, Chiara},
  journal={IEEE Open Journal of the Computer Society}, 
  title={Synthetic Data for Fairness: Bias Mitigation in Facial Attribute Recognition}, 
  year={2025},
  volume={6},
  number={},
  pages={1703-1714},
  keywords={Face recognition;Facial features;Synthetic data;Computer architecture;Annotations;Prevention and mitigation;Accuracy;Training;Image color analysis;Hair;Facial attribute recognition;bias mitigation;demographic fairness;synthetic data;stable diffusion;generative adversarial networks},
  doi={10.1109/OJCS.2025.3622694}}

This project is licensed under the CC BY-NC-SA 4.0 License

You may share and adapt this work for non-commercial purposes only, as long as you give appropriate credit and distribute your contributions under the same license.
For commercial use, explicit permission from the authors is required.