Year: 2025
Type: dissertation
Abstract: Deepfakes present a growing threat to digital media authenticity and cybersecurity, with existing detection models often struggling to generalise across datasets or withstand real-world distortions. This study develops a hybrid Deepfake detection pipeline that combines Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) to better capture both local visual artefacts and global structural inconsistencies. A complete preprocessing workflow was implemented, including video frame extraction, face detection using MTCNN, and image normalisation. The proposed model fuses EfficientNet-B0 and ViT-B/16 feature embeddings to perform frame-level classification, with video-level predictions derived from aggregated softmax probabilities. Evaluation on real video footage demonstrates that the hybrid architecture is functional, reproducible, and capable of producing stable and interpretable Deepfake likelihood scores despite not being trained on Deepfake-specific datasets. The study highlights the complementary strengths of CNN and transformer-based features and identifies key limitations such as lack of temporal modelling and dataset-specific training. The findings provide a foundation for future research into multimodal, temporally-aware, and more robust Deepfake detection systems.
Source: NORMA
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7210164859
