Deepfake Detection Using Multiple Feature Fusion.

It is becoming increasingly easy to generate forged images and videos using pre-trained deepfake methods. The forged images and videos are difficult to distinguish with the human eye, posing security and privacy threats. This chapter describes a deepfake detection method that employs multiple feature fusion to identify forged images. An image is preprocessed to extract high-frequency features in the spatial and frequency domains. Two autoencoder networks are then used to extract the two features of the preprocessed image, following which the extracted features are fused. Finally, the fused feature is input to a classifier comprising three fully-connected layers and a max-pooling layer. The experimental results demonstrate that fusing two features yields an average detection accuracy of 98.62%, much better than using the features individually. The results also show that the method has superior accuracy and generalization capability than state-of-the-art deepfake detection methods.