---
license: mit
task_categories:
- image-classification
language:
- en
---

# MUFAC (Machine Unlearning for Facial Age Classifier)

[![arXiv](https://img.shields.io/badge/arXiv-2311.02240-b31b1b.svg)](https://arxiv.org/abs/2311.02240)

This repository provides a **cleaned and resolution-aligned (128x128) version** of the MUFAC benchmark dataset.

---

## 📘 Description

A multi-class age classification dataset featuring over 86,000 Asian facial images with annotations for age groups and personal identities.

<img src="https://raw.githubusercontent.com/ndb796/MachineUnlearning/main/resources/MUFAC.png" width="700" alt="MUFAC examples"/>

- Preprocessed facial images (128×128 resolution)
- CSV files (`custom_train_dataset.csv`, etc.) for structured training/evaluation
- Separation of identity-forgettable vs. retained images (`forget_images`, `retain_images`)
- Suitable for benchmarking **machine unlearning** algorithms, especially in **task-agnostic setups**

It is specifically intended for experiments where **personal identities are selectively unlearned**, without degrading model utility on the original task (e.g., age classification).

---

## 🗂️ Dataset Structure
```
MUFAC/
├── forget_images/                     # 1,500 images to be unlearned
├── retain_images/                     # 8,525 images to retain
├── train_images_part1/                # 9,999 training images (part 1)
├── train_images_part2/                # 26 training images (part 2)
├── val_images/                        # 1,539 validation images
├── test_images/                       # 4,513 test images
├── fixed_test_dataset_negative/       # 5,000 identity-balanced test data (negative)
├── fixed_test_dataset_positive/       # 5,000 identity-balanced test data (positive)
├── fixed_val_dataset_negative/        # 5,000 identity-balanced val data (negative)
├── fixed_val_dataset_positive/        # 5,000 identity-balanced val data (positive)
├── custom_train_dataset.csv           # CSV with image paths and labels
├── custom_val_dataset.csv             # Validation CSV
└── custom_test_dataset.csv            # Test CSV
```
- CSV files follow the format: `image_path`, `age_group`, `identity`, `forget_flag`, etc.
- All image paths are relative and usable with `datasets.Image()` or `PIL`.

---
## 🔹 How to Use
```python
# Method 1: Git Clone (Recommended)
git lfs install
git clone https://huggingface.co/datasets/Dasool/MUFAC
cd MUFAC
```

# Method 2: Using Hugging Face Hub API
```
from huggingface_hub import snapshot_download

# Download entire dataset
local_dir = snapshot_download("Dasool/MUFAC", repo_type="dataset")
print(f"Dataset downloaded to: {local_dir}")
```
# Method 3: Load CSV and Images
```
import pandas as pd
import os
from PIL import Image

# Load CSV
df = pd.read_csv("MUFAC/custom_train_dataset.csv")
print(f"Dataset size: {len(df)} samples")
print(df.head())

# Load sample image
sample_row = df.iloc[0]
img_path = os.path.join("MUFAC", sample_row["image_path"])
img = Image.open(img_path)
img.show()
```

---
## 📊 Citation

This dataset is part of the benchmark suite introduced in the following paper:

```bibtex
@misc{choi2023machine,
  title={Towards Machine Unlearning Benchmarks: Forgetting the Personal Identities in Facial Recognition Systems},
  author={Dasol Choi and Dongbin Na},
  year={2023},
  eprint={2311.02240},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}
```

- 🔗 Code for Experiments:: [https://github.com/ndb796/MachineUnlearning](https://github.com/ndb796/MachineUnlearning)

- 📧 Contact: dasolchoi@yonsei.ac.kr