An Optimised Deepfake Detection System Using Vision Transformers and FFT

Sabu, Nobin (2025) An Optimised Deepfake Detection System Using Vision Transformers and FFT. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (803kB) | Preview

Abstract

The proliferation of deepfake media—AI-generated synthetic videos that closely resemble authentic content—poses a significant challenge to digital trust, public safety, and content verification. Conventional deepfake detection systems, especially the ones that employ convolutional neural networks (CNNs), tend to fail to perform due to real-world factors, including video compression and non-obvious ways of manipulating them. This study is undertaken to curb such limitations, as it follows in the proposal of a HybridSwinCNN, which is a dual-stream design with advantages of merging a spatial and frequency domain to perform better in terms of accuracy and robustness. The network comprises Swin Transformer to capture the semantic spatial features on RGB pictures and EfficientNet CNN to capture the high-frequency outliers in the log-magnitude spectrum Fast Fourier Transform (FFT).

The study uses the FaceForensics++ data which has 23,988 video frames extracted, preprocessed, trained, and validated using 20 epochs of model. This is evidenced in the fact that the HybridSwinCNN suggested attained the best validation performance with an accuracy of 85.86 percent and a final train accuracy of 99.47 percent, both indicative of good feature learning and generalization capability. The findings indicate the usefulness of the joint use of spatial and frequency features in detecting deep fake and specifically when the material under scrutiny is compressions and adversarially corrupt content. The solution proposed in the work is scalable and interpretable that can be applied in the situation of digital forensics, media verification and online content systems. Future directions will involve the best practise on generalization to unseen forgeries, explainability and making the model suitable to real-time applications.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Gyamf, Eric UNSPECIFIED
Subjects:	Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence > Computer vision Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence > Computer vision
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	26 Aug 2026 11:02
Last Modified:	26 Aug 2026 11:02
URI:	https://norma.ncirl.ie/id/eprint/9661

Actions (login required)

View Item