Deepfake detection using neural networks

Abstract

The widespread use of deepfake technology, which uses advanced artificial intelligence techniques like Generative Adversarial Networks (GANs), poses severe threats to public confidence and digital security. This study investigates sophisticated methods for identifying deepfake content, emphasising the Convolutional Neural Networks (CNNs) ensemble, namely the ResNeXt and EfficientNetB4 models. Using the Celeb-DF (v2) dataset, which includes more than 5,600 deepfake videos, we provide a unique method that combines Siamese training techniques and attentional mechanisms inside CNN structures. With our improvements, EfficientNetB4 could identify edited films with an Area Under Curve (AUC) score of 99% and an exceptional accuracy of 97%. These findings highlight the potential of improved CNN models to differentiate between complex deepfakes and authentic videos accurately. Future research directions include expanding dataset diversity and applying transfer learning to refine detection techniques further, thereby contributing to the secure dissemination of digital content.