Manipulation-equivariant evidence-state learning for multimodal fake news detection

Abstract

Multimodal fake news is often locally plausible but relationally fractured. The image may be authentic, the text fluent, and the named entities recognizable, yet the evidence chain connecting entity, context, visual content, and claim can be broken. Most multimodal detectors encode the observed post as a static image-text state. MEEL (Manipulation-Equivariant Evidence Learning) instead learns typed directions of manipulation-induced evidence-state movement under a fixed-evidence setting. The framework constructs controlled edited views through image swapping, entity shifting, context dropping, relation misbinding, and text overclaiming. Operation-specific displacement prototypes make each manipulation correspond to a predictable direction in evidence-state space. At inference time, no edited view is generated; veracity is predicted from the natural evidence state, its alignment with the learned manipulation directions, sparse direction profiles, and equivariance uncertainty. On MR2-Chinese, MR2-English, and Weibo, MEEL achieves F1/MCC scores of 96.25%/0.9275, 95.21%/0.9231, and 92.25%/0.8296, respectively. Stability analysis, ablations, and direction diagnostics identify semantic edit construction and equivariant direction learning as the main sources of the displacement signal; scalar cues, pseudo-news supervision, random-edit controls, and uncertainty modeling have auxiliary or diagnostic roles. The model learns the movement of evidence states under semantic manipulation, beyond static image-text compatibility.