Evaluation of Convolutional Neural Networks Against Deepfakes Using Transfer Learning
View/ Open
Abstract
The main objective of this paper is to evaluate ResNets, DenseNet, Inception and VGG, against deepfake images, to answer the question: How effectively these Convolutional Neural Network can distinguish between deepfake images and real images.
The dataset was acquired from FaceForensics++ and CelebA datasets for manipulated and unmanipulated images respectively. A custom script using Python and OpenCV was applied to create the final dataset for modelling.
Transfer learning is a technique of applying the learned features by a network to a similar approach. It is employed to save time and resources in training, as it does not require a large dataset to allow the network to learn effectively.
The Convolutional Neural Networks are tested against different deep fakes and the networks are evaluated using metrics like precision, recall, accuracy, loss, and f-1 score. It was observed that all the networks used in the experiment performed exceptionally well, but Inception network was slightly better than the other networks in separating the real and fake images.