DeepfakeTIMIT Deep Fake Detection Dataset

* This dataset supports online use.Click here to jump.

DeepfakeTIMIT is a dataset for deepfake detection created by Idiap Research Institute in 2018. It contains videos of faces swapped using an open source generative adversarial network (GAN)-based method. These videos are created based on the original autoencoder-based Deepfake algorithm. The dataset manually selected 16 pairs of individuals with similar faces from the publicly available VidTIMIT database, and trained two models of different qualities for each individual: a low-quality (LQ) model with an input/output size of 64×64, and a high-quality (HQ) model with a size of 128×128. For each person's 10 videos in the VidTIMIT database, 320 corresponding versions of the video were generated, resulting in a total of 620 videos with swapped faces. In terms of audio, the original audio track of each video was retained, and no modifications were made to the audio channel.

This dataset is designed to support research on deep fake detection techniques and can be used to train and test related deep learning models. The relevant reference paper is the paper by P. Korshunov and S. Marcel.DeepFakes: a New Threat to Face Recognition? Assessment and Detection”, and C. Sanderson and BC Lovell’s paper “Multi-Region Probabilistic Histograms for Robust and Scalable Identity Inference》.