Reproduce “The Coherence Trap: When MLLM-Crafted Narratives Exploit Manipulated Visual Contexts” on OpenResearch

About this paper

Yuchen Zhang, Yaxiong Wang, Yujiao Wu, Lianwei Wu, Li Zhu, Zhedong Zheng

This research addresses the detection of sophisticated multimedia misinformation generated by Multimodal Large Language Models (MLLMs), where manipulated visuals are paired with semantically coherent, fabricated narratives. It introduces the MDSM dataset, featuring 441,423 samples with MLLM-generated high-fidelity texts, and the AMD framework, which achieved an average performance of 88.18 ACC, 60.25 mAP, and 61.02 mIoU in cross-domain detection while being more efficient than existing MLLM-based methods.

Key insights

• MLLM-driven deception, involving manipulated visuals paired with semantically coherent fabricated narratives, presents a more sophisticated and challenging threat to information integrity than previously studied forms of misinformation.
• MLLM-based detection methods leverage inherent world knowledge, allowing them to achieve superior performance compared to forgery-trace methods, particularly in scenarios where semantic alignment between manipulated visuals and text is maintained.
• Text fabrications generated by MLLMs are notably more challenging to detect than image-modal manipulations, underscoring the advanced deceptive capabilities of MLLMs in crafting nuanced linguistic content.

Summary by alphaXiv.