Implementation of Deep Learning Techniques for Deepfake Classification: A comparative study using ResNet-50 and VGG16

This work investigates the use and assessment of ResNet-50 and VGG16 deep learning models for detecting deepfake images using the Labelled Faces in the Wild (LFW) dataset. The methodology encompassed thorough preparation of the dataset and the utilisation of both models to differentiate between authentic and altered facial photos. The models' performance was evaluated using fundamental criteria including accuracy, precision, recall, and F1 score. VGG16 demonstrated superior performance compared to ResNet-50, achieving an accuracy of 93.87%, recall of 96.23%, and an F1 score of 94.41%, while ResNet-50 achieved a precision of 92.54%, recall of 95.26%, and F1 score of 93.88%. Although both models achieved good accuracy, VGG16 had superior stability and generalization across the whole dataset. The study determines that VGG16 is the more dependable model for detecting deepfakes on the LFW dataset. It suggests that future efforts should concentrate on enhancing ResNet-50 or creating ensemble methods to enhance detection capabilities.