Deepfake Video Detection Using Spatiotemporal Convolutional Network and Photo Response Non Uniformity

It is important to improve the detection of deepfake videos to differentiate between real and fake videos that cause disinformation in the digital age so that a high level of accuracy is required. The purpose of deepfake video detection is to aid digital content consumers to surmount disinformation and sever real videos from fake ones. Limited by the number and quality of datasets, the time required for detection, and consistent performance evaluation i.e., the detection model cannot detect videos detected with video editing tools. This study provides a solution to this problem by using the Spatiotemporal Convolutional Network (SCN) method and Photo-Response Non-Uniformity (PRNU) analysis. The dataset used will go through pre-processing stages, extract per-frame video, detect face parts, and face cropping. Then the data is spread and modeled using RestNext50 and LSTM. This study produced 10 models using the FaceForensic, CelebDF, and DFDC datasets, and a mixture of these datasets which can then be used to analyze deepfake videos. The test results show that the deepfake detection process is faster and more accurate with an accuracy rate of up to 97.89%.

I. Introduction

Deepfake video detection is a method to extract and classify real video or fake video made using artificial intelligence [1] [2]. Deepfake video provokes distortion and anxiety due to its potential usage for maleficent objectives such as defamation, political bias, sabotage, intimidation, exploitation, false propaganda, piracy, and other types of notorious activity [3] [4]. The availability of deepfake detection methods can aid digital content consumers to surmount disinformation and sever real videos from fake ones [5] [6]. Nevertheless, technological progress especially in machine learning-based software has hastened and eased the process to make deepfake video appears real though leaving some manipulation traces [7] [8]. Alas, the deepfake detection method perfection is hindered by dataset quantity and quality, detection process, and performance evaluation inconsistency causing the detection model's inability to recognize manipulated video made using video editing tools [7]. Therefore, a technique to detect deepfake video is required.