Explainable Deepfake Detection: Leveraging Vision-Language Models for Multimedia Authentication

23^rd April 2026, 11:00 Ashton 208
Guangliang Cheng
UoL

Abstract

This presentation addresses the pressing challenge of multimedia authenticity in the era of generative AI, and explores how vision-language models can drive deepfake detection beyond binary classification toward interpretable, localizable, and verifiable authentication. Covering images, videos, and unified multimodal social media content, the talk presents our recent advances in large-scale benchmark construction, forgery detection and localization, cross-modal explanation, and deployable demonstration systems, including SIDA, So-Fake, BusterX, and Omni-Fake. By moving from detection to trustworthy understanding, this presentation highlights a broader vision for building transparent and reliable AI-powered infrastructures for the future digital society.