DeepFake Detection Challenge Train Set V2

The DeepFake Detection Challenge (DFDC) Train Set V2 is a massive, multi-modal dataset developed by Meta AI (formerly Facebook) in collaboration with AWS, Microsoft, and academic partners. Engineered to combat the proliferation of highly realistic synthetic media, the dataset consists of over 124,000 videos across a diverse range of 3,450 actors. Architecturally, the dataset is designed to provide high-fidelity 'ground truth' for both real and manipulated content, utilizing various generative adversarial network (GAN) techniques and face-swapping algorithms. As of 2026, it remains the baseline for benchmarking deepfake detection robustness due to its inclusion of complex lighting conditions, varied backgrounds, and diverse human subjects, which significantly reduces demographic bias in resulting models. The technical architecture of the dataset allows for the training of advanced Vision Transformers (ViTs) and Convolutional Neural Networks (CNNs) capable of detecting temporal inconsistencies and frequency-domain artifacts that characterize 2026-era synthetic media. It is distributed under the DFDC agreement, facilitating non-commercial research and development for the global cybersecurity community.