VidAvDetect: A Deepfake-Inspired Vision Transformer Approach for Detecting Real Humans vs. AI-Avatars in Video Streams

The pace of advancement in Generative AI has made it possible to realize highly realistic synthetic identities in the form of avatars for non-existent persons, thus paving the way for a paradigm beyond state-of-the-art deepfake attacks that aim to manipulate real identities in people. This rapidly emerging trend poses a challenge to digital media forensics in a most critical way, in terms of deciding whether a facial identity observed in a video clip represents a real human identity versus a fully synthetic identity created using advanced tools in the realm of Generative AI. To address this gap, we introduce VidAvDetect, a deepfake-inspired Vision Transformer approach specifically designed to discriminate real human faces from AI-generated avatars in video streams, addressing a novel identity-existence verification task. The proposed system integrates efficient frame sampling, robust facial preprocessing, patch-based embeddings, and global structural modeling through a transformer encoder, enabling the detection of subtle geometric and textural regularities characteristic of synthetic identities. Experimental results demonstrate strong performance, with training accuracy reaching 97–98%, video-level accuracy of 95.1%, a macro F1-score of 0.944, and a ROC-AUC of 0.991, confirming the model’s robustness across heterogeneous real, manipulated, and fully synthetic datasets. By moving beyond manipulation detection to focus on identity-existence verification, VidAvDetect establishes a new methodological direction for transparency, regulation, and trust in modern digital media environments where AI-generated avatars increasingly resemble real humans.

@article{Acim2025,
  title     = {VidAvDetect: A Deepfake-Inspired Vision Transformer Approach for Detecting Real Humans vs. AI-Avatars in Video Streams},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {16},
  number    = {12},
  year      = {2025},
  publisher = {The Science and Information Organization},
  author    = {Btissam Acim and Hamid Ouhnni and Nassim Kharmoum and Soumia Ziti},
  doi       = {10.14569/IJACSA.2025.01612110},
  url       = {https://doi.org/10.14569/IJACSA.2025.01612110}
}