Speakers
Description
In the era of rapidly advancing technology, the spread of deepfake content has become an escalating challenge, making detecting synthetically created materials crucial in softening the potential harms. This paper focuses on a comparative analysis of Convolutional Neural Networks (CNNs) designed explicitly for deepfake detection (MesoNet and MesoNet Inception) with a custom architecture proposed by the authors, demonstrating a high degree of precision in distinguishing between authentic and manipulated content. The tested models consisted of a few layers, which enabled focusing on mesoscopic features in the images. The networks were trained on a dataset generated using the CelebA dataset and FaceDancer, a face-swapping method. Additionally, the networks were tested on additional deepfakes to evaluate their generalization ability. This study critically evaluates the effectiveness of the solutions and the impact of dataset selection.
