Fake-Vs-Real-Faces (Hard)

This dataset is designed to challenge and advance the capabilities of machine learning models in classifying real and fake human faces. Unlike most available datasets that lack variance, this dataset ensures a diverse range of features for real faces and uses StyleGAN2-generated fake faces, creating a more difficult classification task, even for the human eye.

Key Features

✅ Challenging Classification Task: Includes high-quality, realistic fake faces generated by StyleGAN2, raising the difficulty for accurate classification.
✅ Diverse Real Faces: Real human faces are gathered with attention to age, sex, ethnicity, makeup, and other features, ensuring a realistic and fair representation for production scenarios.
✅ Uniform Image Format: All images are in JPEG format with a consistent resolution of 300×300 pixels for seamless AI training.
✅ Balanced Dataset: Contains a total of 1,288 images:

• Fake Faces: 700 images (collected from thispersondoesnotexist.com).
• Real Faces: 589 images (extracted via the Unsplash API and processed using OpenCV).

Dataset Contents

• Images: 1,288 labeled JPEG images.
• Labels: A data.csv file mapping image IDs to their respective labels (Fake or Real).

Applications

• Deep Learning Model Development: Train neural networks for face classification tasks.
• AI for Fraud Detection: Detect GAN-generated fake faces in real-world applications.
• Facial Recognition Research: Enhance the robustness of AI systems in handling diverse features and synthetic face detection.