Where Detectors Fail: Closing the Tail-Domain Gap with Expert-Guided Mutual Distillation

2026-07-29

AI Summary

Existing multimodal fake news detectors exhibit limited generalization in cross-domain scenarios, often misled by unreliable cues such as domain shortcuts induced by data imbalance and semantic inconsistencies between images and text. To address this, this work proposes an Expert-Guided Mutual Distillation (EGMD) framework, which establishes a three-level cooperative trust evidence learning mechanism spanning input, representation, and decision stages. At the input level, it calibrates image-text consistency; expert-guided teachers align domain statistics and emphasize domain-specific patterns, while prototype-anchored student models inherit the teachers’ feature geometry and calibrated predictions through mutual learning and dual-channel distillation. Experiments show that EGMD achieves state-of-the-art performance across four cross-domain datasets and reduces domain bias by up to 57.3%. Additionally, we introduce Weibo_Balanced, the first domain-balanced benchmark, to disentangle the impact of data imbalance on model generalization.

Abstract

Multimodal fake news detectors often generalize poorly across domains because they learn to trust unreliable evidence: domain-specific shortcuts amplified by imbalanced data and semantically inconsistent text-image pairs that make cross-modal evidence unreliable. We propose Expert-Guided Mutual Distillation (EGMD), which learns what evidence to trust across the prediction pipeline. At the input level, input-level calibration encodes pair-level coherence as a shared gain before fusion. At the representation level, an expert-guided teacher aligns domain statistics and encourages domain-specific patterns to concentrate in specialized experts. At the decision level, prototype-anchored domain-specific students use mutual learning and dual-channel distillation to inherit the teacher's feature geometry and calibrated predictions while discouraging local domain priors. We further construct Weibo_Balanced, a domain-balanced benchmark that isolates the effect of imbalance on generalization. Across four datasets in two languages, EGMD achieves state-of-the-art accuracy while reducing domain bias by up to 57.3%.

Problem

• multimodal fake news detection
• domain generalization
• data imbalance
• cross-modal inconsistency
• domain-specific shortcuts

Innovation

• Expert-Guided Mutual Distillation
• Multimodal Fake News Detection
• Domain Generalization
• Mutual Learning
• Calibrated Prediction