Alam Rahmatulloh, Herman Dwi Surjono, Fatchul Arifin, Irfan Darmawan, Nia Ambarsari
The increasingly massive proliferation of deepfake content poses a serious threat to the authenticity and trustworthiness of digital information. This study proposes VERITAS, a deepfake detection framework that integrates CNN-based feature extraction, Vision Transformer (ViT) architecture, and multilevel Squeeze-and-Excitation (SE) blocks to adaptively strengthen spatial and token attention. To enhance generalization across datasets, VERITAS also adopts a dual-branch self-learning mechanism consisting of Masked Image Modeling (MIM) and Identity Distillation (ID) based on the Face Security Foundation Model (FSFM) framework. Experimental results show that VERITAS achieves AUC scores of 87.45% and 88.30% on the Celeb-DF v2 and WildDeepfake datasets at frame-level, and 95.56% and 87.75% at video-level, respectively, outperforming various state-of-the-art methods. Ablation studies confirm the significant contributions of each component of the architecture. Despite its strong detection performance, inference speed remains a challenge, so further research directions include optimizing the model for real-time deployment. This research contributes to building a robust and adaptive deepfake detection system across domains. © This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License. License details: https://creativecommons.org/licenses/by-sa/4.0/
Faculty of Engineering, Universitas Negeri Yogyakarta, Yogyakarta, Indonesia; Faculty of Engineering, Universitas Siliwangi, Tasikmalaya, Indonesia; Faculty of Industrial Engineering, Telkom University, Bandung, Indonesia
