Preprocessing and Feature Extraction based Deepfake Detection on Combined dataset

There are many deepfake detection models available, each with their own limitations. This research concentrates on one such limitation and addresses it. That research gap would be the inefficient behavior of the deepfake detection model on the orientation of faces in the input image or video. The architecture starts by combining the data sets of DFDC, Celeb-DF, and Face Forensics++. Then frame-extraction and pre-processing on the frames take place. The next stage is crucial and differs from other models; it is the feature extraction stage, where specific face-orientation-related features are extracted from each frame. The features that are extracted are a histogram of oriented gradients, local binary patterns, and scale-invariant feature transforms. Later on in the architecture, normalization of features happens in independent neural network pipelines. Then features are combined and passed onto the next neural network, which will classify the videos as deepfake or not. The validation accuracy of this architecture is 95.89%.