A Comparative Review of Deep-Learning Models for Deepfakes Detection
Résumé
The development of generative AI has advanced significantly over the past few decades, enabling the creation of deepfake images that are increasingly difficult to distinguish from genuine photographs. The widespread availability of these models, which can be easily used, poses a substantial risk of spreading disinformation. Consequently, there is a pressing need for robust and reliable methods to identify images that have been created or altered using generative AI models. To address this need, a diverse range of methods and models have been developed. However, these approaches are often not exhaustively compared to one another, nor are they evaluated using a common reference dataset. Moreover, the majority of existing deepfake detection models rely on deep learning techniques, with numerous models available for feature extraction and detection, ranging from simple Convolutional Neural Networks (CNNs) to more advanced Vision Transformers (ViTs). To ensure the comparability and reproducibility of deepfake detection models, a standardized benchmark is urgently required to evaluate their performance across a large-scale, common dataset. Such a reference benchmark would facilitate the development of more effective and robust detection methods by providing insight into the strengths and weaknesses of existing AI-based approaches. In addition to establishing this benchmark, this paper also explores the challenges of combining different AI-based deepfake detection models and investigates various aggregation methods to further improve overall detection performance. A large-scale experiment involving almost 50 generative AI methods and over 40 deep learning-based feature extraction and detection models demonstrates the relevance of this study.
Origine | Fichiers produits par l'(les) auteur(s) |
---|