This research presents a novel paradigm that transforms vision-language models (VLMs) into powerful tools for identifying manipulated media with improved generalizability and explainability.

• Introduces a knowledge-guided forgery adaptation module that aligns VLMs' semantic understanding with forensic features
• Leverages contrastive learning to enhance detection capabilities
• Develops a more generalizable approach to identifying deepfakes across various manipulation types
• Provides explainable results that help users understand why content is flagged as fake

In an era of increasing digital misinformation, this advancement offers critical security benefits by improving verification of content authenticity and protecting against sophisticated media manipulation techniques.

Unlocking the Capabilities of Vision-Language Models for Generalizable and Explainable Deepfake Detection