Spatial Feature-based Fake Capsule Network Model for Deep fake Detection for Image and Video Data

Abstract

The development in the image or video editing techniques paved the way for attackers to make fake videos and images. To overcome this problem, Convolutional Neural Network (CNN) techniques had been introduced and it had delivered substantial results. But the fake videos created using the Deepfake tool had been challenging to the existing CNN techniques. Also, CNN has drawbacks such as the network being significantly slow due to max pool operation and requiring a large dataset to train and process the neural network. The drawbacks of CNN can be overcome by Capsule Networks to detect Deepfake videos. Spatial Feature based Fake Capsule Network Model (FCNM) is proposed to detect fake news through images and video. The FCNM model comprises of Capsule structures, Exponential Linear Unit (ELU), LP Pooling layer and dynamic routing algorithm. The detection performance of the proposed Capsule Network over the attacks such as Deepfake, Face2face and FaceSwap had produced significant results.