Foundation model for detecting AI-generated visual content

Thumbnail Image

Files

master_Klechshev_Maxim_2025.pdf (39.81 MB)

URL

Journal Title

Journal ISSN

Volume Title

School of Science | Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.

Authors

Date

2025-11-21

Department

Major/Subject

Machine Learning, Data Science and Artificial Intelligence

Mcode

Degree programme

Master's Programme in Computer, Communication and Information Sciences
Master's Programme in Computer, Communication and Information Sciences
Master's Programme in Computer, Communication and Information Sciences

Language

en

Pages

81

Series

Abstract

The vast quantity of synthetic content produced by modern, accessible generative AI models and the lack of robust detection tools presents a significant danger to information integrity across the internet. This study intoduced a comprehensive foundation model approach for deepfake detection, demonstrating a strong performance across a diverse range of generative models. Following the philosophy of foundation model development, we document the complete architectural evolution: from traditional CNN baselines through multi-modal ensembles to our final PE-Giant adaptation, offering a systematic roadmap for the subject. Trained on a self-generated 5.2M-sample synthetic dataset with text-to-image, image-to-image, text-to-video and image-to-video generation methods, our model achieves state-of-the-art results in distinguishing authentic, fully-synthetic and partially-synthetic content. Most importantly, the model demonstrates robust cross-domain generalization maintaining high accuracy not only on unseen data by the same architectural family models but also on content generated via entirely separate synthesis paradigms. We also organised a big 48 hours AI hackathon in Finland to bring attention to the problem.

Description

Supervisor

Marttinen, Pekka

Thesis advisor

Ilin, Alexander

Keywords

deepfake detection, AI-generated content, foundation models, inpainted images, diffusion models, computer vision

Other note

Citation

Permanent link to this item

https://urn.fi/URN:NBN:fi:aalto-202512179379

Collections

[dipl] Perustieteiden korkeakoulu / SCI