Recognition of Deepfakes Generated Through Al
Year: 2024
Type: article
Abstract: This work presents a research trajectory dedicated to the analysis and development of Artificial Intelligence models for recognising deepfakes, with a particular focus on human faces in images and videos. The study examines state-of-the-art approaches in computer vision and machine learning, moving progressively from foundational experiments to advanced detection systems. Throughout the research, several tools and methods were designed, including automated procedures for extracting face frames from video datasets and integrated architectures that combine ResNet-50 feature extraction with LSTM-based temporal modelling. These systems were tested on custom datasets, achieving high performance, including an AUC of 0.98, accompanied by detailed evaluations through confusion matrices and error analyses. The project also incorporates semi-supervised labelling methods, developed in collaboration with the Universidad de Cádiz, to enrich the FaceForensics++ dataset with additional facial attributes. Furthermore, it investigates the generalisation problem in deepfake detection by examining model behaviour across different manipulation techniques and by applying Grad-CAM to understand the visual patterns guiding classification. The results highlight both the progress achieved and the remaining challenges, particularly regarding overfitting, the reduction of false positives, and the need for improved robustness across heterogeneous deepfake scenarios. Overall, the work contributes methodological tools, practical insights, and research directions that support the development of more reliable and interpretable deepfake detection systems.
Author: Stile, Vittorio
Institution: Mercatorum University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7115166390
FWCI:
Cites:
Cited by:
Related to: