DeepFake Detection Using Convolutional Neural Networks

Abstract

DeepFakes, or facial modifications produced by artificial intelligence, pose serious threats to the integrity of digital content and public trust. This paper proposes a method founded on Convolutional Neural Networks (CNNs) for binary facial image classification, separating real from fake. A properly processed and prebalanced dataset is utilized, and multiple data augmentation techniques are applied to ensure the model's maximum generalization capability. The proposed CNN structure, although relatively simple, is a sequence of several convolutional and pooling layers followed by dense layers with dropout regularization applied. Model training is performed with binary cross-entropy loss and optimized using the Adam optimizer. The experiment outcomes indicate that the model performs effectively on unseen test samples, thus demonstrating its capability to detect manipulated facial images and thus contributing to the overall goal of media integrity assurance.