Ensemble of optimal feature descriptors for Deepfake video detection

Deepfake videos, employing artificial intelligence techniques to create highly realistic but fabricated content, have emerged as a major concern for the society. Proposed paper develops deep learning model called, EnsembleResnet, that ensembles multiple feature descriptors, and then judicially select optimal features for deepfake video detection. Initially, individual features such as DFT (Discrete Fourier Transform), DCT (Discrete Cosine Transform), SIFT (Scale-Invariant Feature Transform), and Gabor Filter are trained on ResNet-18 model to classify real and fake videos. Afterward, ensemble learning of optimal feature descriptors is trained on the same model to improve detection accuracy. Experimental evaluations on diverse datasets: FF++ and Celeb-DF show potency of ensemble learning in identifying real and fake videos. Experimental results show an accuracy with paired feature descriptors (SIFT + Gabor + DCT) on datasets FF++ (raw data, high quality, low quality) is (99.80%, 99.25%, 99.79%), and on Celeb-DF dataset is 98.01%, respectively.