MMFakeBench is the first comprehensive benchmark tackling the challenge of detecting misinformation that combines multiple types of manipulations across text and images.

• Addresses a critical gap in misinformation detection by handling mixed-source forgeries
• Covers three key distortion types: textual veracity, visual veracity, and text-image consistency
• Enables more robust evaluation of Large Vision-Language Models for detecting complex, real-world misinformation
• Sets a new standard for security tools that can identify sophisticated multimedia fakes

This research is essential for developing defense mechanisms against increasingly sophisticated misinformation campaigns that combine multiple manipulation techniques to appear more credible.

MMFakeBench: A Mixed-Source Multimodal Misinformation Detection Benchmark for LVLMs