Deep Learning–Based Fake Image Detection Using Transfer Learning: A Comprehensive Review

Abstract

The fast evolution of artificial intelligence has transformed the development of the digital content, allowing to create very realistic synthetic images and videos. Although these technologies have many legal uses, they have also given rise to the development of false visual content or deepfakes, which is highly dangerous to the integrity of information, social trust, computer security, and even investigations. Detecting these manipulated images has now become a crucial task and in some cases even standard algorithms based on handcrafted features has been found to be ineffective against advanced generative algorithms such as Generative Adversarial Networks (GANs) and diffusion-based algorithms. Transfer learning Deep learning-based methods, and in particular those that employ techniques of transfer learning, have become valuable solutions because they can extract discriminative features on small datasets at a lower cost. The given paper provides the in-depth discussion of the existing methods of fake image detection based on deep learning and transfer learning. It discusses popular ready-made convolutional neural network architectures, test datasets, metrics, and the current trends in research. Comparative studies point out the advantages and disadvantages of the current tools, and the paper also reveals key problems and sketches the way forward in creating robust, scalable and generalizable fake image detector systems that could help counter the emerging challenges in cyberspace.