Smartphone-based Deepfake Detection Through Transfer Learning and Lightweight Knowledge Distillation Technique

The advancement of artificial intelligence technology has been growing rapidly. Generative adversarial network, a subfield of artificial intelligence technology, generates fake faces with more realistic features, which are much harder to identify by human beings. Deepfake is the concept used to describe the creation or alteration of films, images, or audio using artificial intelligence or machine learning methods to mislead or deceive audiences. In this study, we proposed an image classification model that can classify deepfake images with a higher level of accuracy. The developed model in this research attempts to use a custom CNN model, Knowledge Distillation technique, and pre-trained models. Pre-trained models are used to avoid the computational cost of training a CNN from scratch. The images of deepfake employed in this work were fed into the deep convolutional neural network without extensive preprocessing, and the deep features from the images were extracted in several layers. The hyperparameters of the custom CNN model have been optimized using the cuckoo search optimization (CSO)-based metaheuristic technique. The Xception and DenseNet models achieved the best performances with test accuracies of 99% and 98%, respectively. The mean training time of Knowledge Distillation is 115 seconds for every epoch, significantly lower than other models. The Knowledge Distillation model has been utilized to expand the deepfake detection system into a website and Android smartphone application due to its small number of parameters and lowest training time. Finally, the developed mobile application was evaluated by 50 volunteers.