Abstract

Digital authentication systems that rely on biometric recognition are especially vulnerable to deepfake attacks, which can be used to impersonate legitimate users and bypass security protocols. As deepfake attacks become increasingly sophisticated, detection methods must evolve rapidly. In this paper, we propose the usage of transfer learning instead of standard deep learning to provide a fast response to novel threats. We evaluate 12 approaches, combining three deep neural networks as feature extractors with four traditional machine learning algorithms as classifiers. Finally, the best-performing model, i.e. ConvNeXt with a support vector classifier, is fine-tuned and evaluated on a real-world dataset, demonstrating strong performance.

Recommended Citation

Papić, M., Milošević, P., Milenković, I., Milovanović, M. & Minović, M. (2025). Transfer Learning for Deepfake Detection in Static Facial ImagesIn I. Luković, S. Bjeladinović, B. Delibašić, D. Barać, N. Iivari, E. Insfran, M. Lang, H. Linger, & C. Schneider (Eds.), Empowering the Interdisciplinary Role of ISD in Addressing Contemporary Issues in Digital Transformation: How Data Science and Generative AI Contributes to ISD (ISD2025 Proceedings). Belgrade, Serbia: University of Gdańsk, Department of Business Informatics & University of Belgrade, Faculty of Organizational Sciences. ISBN: 978-83-972632-1-5. https://doi.org/10.62036/ISD.2025.81

Paper Type

Poster

DOI

10.62036/ISD.2025.81

Download

Share

COinS
 

Transfer Learning for Deepfake Detection in Static Facial Images

Digital authentication systems that rely on biometric recognition are especially vulnerable to deepfake attacks, which can be used to impersonate legitimate users and bypass security protocols. As deepfake attacks become increasingly sophisticated, detection methods must evolve rapidly. In this paper, we propose the usage of transfer learning instead of standard deep learning to provide a fast response to novel threats. We evaluate 12 approaches, combining three deep neural networks as feature extractors with four traditional machine learning algorithms as classifiers. Finally, the best-performing model, i.e. ConvNeXt with a support vector classifier, is fine-tuned and evaluated on a real-world dataset, demonstrating strong performance.