COMPARATIVE EVALUATION OF DEEPFAKE DETECTION MODELS FOR SECURE REMOTE BIOMETRIC VERIFICATION
TL;DR
Results indicate that hybrid convolutional-transformer architectures generalize to boundary-free synthesis in a way that purely convolutional detectors do not, and that an open-source model can approach commercial detection quality without transmitting biometric data to a third party.
Abstract
This study examines how remote biometric verification systems hold up against next-generation video deepfakes. Financial and digital platforms increasingly use video-based onboarding, but these pipelines are vulnerable to image-to-video animation, which generates video from a single static image without leaving the spatial splicing or blending artifacts that older detectors look for. To evaluate this threat, we developed a balanced dataset of 51 videos, consisting of 26 authentic videos and 25 synthetic videos generated via face-merging and subsequent image-to-video animation. We benchmarked two commercial SaaS platforms (TruthScan and Hive Moderation) against six open-source deep learning models: EfficientNet-B7, a CLIP-based detector, GANomaly, two Vision Transformers, and GenConViT. Classical convolutional networks such as EfficientNet-B7 did not generalize to this synthesis method: recall fell to 0.240. The hybrid GenConViT model gave the most balanced performance of the open-source systems, with an accuracy of 0.843, an F1-score of 0.825 and a specificity of 0.923. That accuracy is close to the commercial platforms, which scored between 0.882 and 0.901. These results indicate that hybrid convolutional-transformer architectures generalize to boundary-free synthesis in a way that purely convolutional detectors do not, and that an open-source model can approach commercial detection quality without transmitting biometric data to a third party.