Data for RealFake-450K: A Balanced Dataset of Authentic and Manipulated Images for Deepfake Detection

Abstract

This research introduces a large-scale deepfake image dataset designed to support the development and evaluation of deep learning models for detecting manipulated facial content. With the growing threat of synthetic media, particularly deepfakes, the ability to distinguish real from fake images has become vital for ensuring digital trust and authenticity.

The dataset comprises 451,440 images, evenly split between real and fake samples, making it one of the most comprehensive deepfake image collections to date. It combines data from six reputable sources, including FaceForensics++, Celeb-DF, CDDB, WildDeepfake, DeeperForensics-1.0, and Kaggle’s 140K Real and Fake Faces dataset. These images capture a diverse range of facial identities, lighting conditions, compression levels, and forgery techniques, providing a robust foundation for training models that generalize well in real-world applications.

This dataset was instrumental in developing and benchmarking an enhanced deepfake detection framework for the AI-Guard mobile app. Multiple CNN architectures—including VGG19, InceptionV3, Xception, EfficientNetB0, ResNet50, and MobileNetV3Large—were fine-tuned using this dataset. The best-performing model (VGG19) achieved a validation accuracy of 98.92% and performed reliably on unseen real-world data. Lightweight models like MobileNetV3Large and EfficientNetB0 also showed promising results for mobile deployment.