DeepFakeGuard: A Deep Transfer Learning Framework for Accurate Video Deepfake Detection

The idiom “Seeing is believing” probably doesn't work in this digital age, where technology is becoming increasingly open to manipulation and creation of deepfakes can be made right on a smart phone. These AI-created deepfakes, significantly complicate the identification by the naked eye of real and false information. As such, deepfake detection has emerged as a challenging, impactful problem in many domains. This work presents a recent technique for the detection of video deepfake based on deep learning models InceptionResNextV2 and Long Short-Term Memory (LSTM). Utilizing transfer learning, the pre-trained InceptionResNextV2 Convolutional Neural Network (CNN) is used to extract feature of the video data, while the extracted feature is further processed by the LSTM network for accurate classification. Using an 80:10:10 train–validation–test split, experiments were carried out on the FaceForensics++ (FF++) dataset (c23 and c40 compression levels). The proposed method achieved as high as 98.97% and outperformed the state-of-the-art methods in detecting the deepfake videos. This express the efficacy of the joint LSTM and InceptionResNextV2 model for combating the increasing threat of deepfake in the current era of modern computing.