Transferência de aprendizado em redes neurais convolucionais para detecção de deepfakes faciais.

Abstract

The proliferation of highly realistic synthetic media (deepfakes) demands robust detection methods. This study investigates the effectiveness of Convolutional Neural Networks (VGG16, ResNet-152, and Inception-v3) in identifying facial deepfakes using Transfer Learning. The methodology encompassed feature extraction and fine-tuning experiments on a realistic dataset from the literature. Internal results highlighted the ResNet-152 architecture as the bestperforming model, achieving over 99% accuracy after fine-tuning. However, external validation tests on unseen images revealed a drop in network performance, emphasizing the critical generalization challenge models face against novel synthesis methods. We conclude that while the evaluated architectures are highly effective within known domains, it is essential to develop strategies that enhance their generalization capacity to keep pace with the rapid evolution of generative artifical intelligence technologies.