CORE: Conflict-Oriented Reasoning for General Multimodal Manipulation Detection
Researchers have introduced CORE, a novel framework designed to detect manipulated multimodal content by identifying inherent conflicts. This approach leverages multimodal large language models (MLLMs) to capture semantic or physical inconsistencies across different data types or with general knowledge. To train these models, a new dataset called the Conflict Attribution Corpus (CAC) was created, featuring detailed annotations of conflict factors. CORE demonstrates robust and generalizable detection capabilities, outperforming existing methods even in zero-shot scenarios. AI
IMPACT Introduces a new method for detecting sophisticated AI-generated misinformation, potentially improving trust in digital content.