DEEPFAKE DETECTION BASED ON DEEP LEARNING

Abstract

The spread of deepfake images and videos in cyberspace is a threat to information security area. Besides, in recent years, deep learning has been increasingly developed and widely applied in many fields, and algorithms have improved performance and accuracy. Hence, the application of deep learning in deepfake detection is a practical research direction. However, applying deep learning models requires a lot of data to solve the problem effectively and takes lots of time to perform training. There is a commonly applied method that helps improve accuracy and takes advantage of pre-trained models with good quality and high accuracy called transfer learning. This study introduces an approach to detect deepfake images using transfer learning methods, including XceptionNet, RestNet101, InceptionResV2, MobileNetv2, VGG19 and DenseNet121, along with comparing it with a traditional CNN model. Through experiments on the Celeb-DF dataset, we demonstrate that DenseNet121 and the softmax classifier perform better than the rest of the methods.