A multi-color spatio-temporal approach for detecting DeepFake

Abstract

The current surge in hyper-realistic faces created artificially using DeepFakes necessitates media forensics solutions suited to video streams and perform reliably with a low false alarm rate at the video level. The paper proposes a spatial and temporal aware pipeline to detect DeepFake videos automatically. Our method employed a two-stream convolutional neural network to extract local spatial and temporal features independently. These features are then fed to fully connected layers to classify whether a video has been subject to manipulation. The proposed method has been evaluated against FaceForensics++, DFTIMIT, and DFD benchmarks. Our suggested technique demonstrates encouraging performance in this task.