A Performance Enhancement of Deepfake Video Detection through the use of a Hybrid CNN Deep Learning Model

Abstract

In the current era, many fake videos and images are created with the help of various software and new AI (Artificial Intelligence) technologies, which leave a few hints of manipulation. There are many unethical ways videos can be used to threaten, fight, or create panic among people. It is important to ensure that such methods are not used to create fake videos. An AI-based technique for the synthesis of human images is called Deep Fake. They are created by combining and superimposing existing videos onto the source videos. In this paper, a system is developed that uses a hybrid Convolutional Neural Network (CNN) consisting of InceptionResnet v2 and Xception to extract frame-level features. Experimental analysis is performed using the DFDC deep fake detection challenge on Kaggle. These deep learning-based methods are optimized to increase accuracy and decrease training time by using this dataset for training and testing. We achieved a precision of 0.985, a recall of 0.96, an f1-score of 0.98, and support of 0.968.