This research introduces a new framework for detecting multimodal misinformation, particularly focusing on miscaptioned images that spread false information.

• Develops a latent multimodal reconstruction (LMR) approach that learns to identify mismatches between images and captions
• Addresses the challenge of limited training data by creating synthetic examples
• Significantly outperforms existing methods on benchmark datasets
• Provides practical tools for fact-checkers to identify manipulated content

This work has critical implications for digital security by helping combat the spread of visual misinformation that can influence public opinion and undermine trust in media.

Latent Multimodal Reconstruction for Misinformation Detection