Skip to content
Book Open access

Multimodal Fake News Detection via Gradient Graph-based Label Propagation and Gradient-Guided Denoising Masking

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 51 references

Abstract

The rapid proliferation of social media has accelerated multimodal misinformation, where false claims paired with manipulated images make detection substantially more challenging. Existing multimodal fake news detection methods typically predict each sample independently, ignoring inter-sample relational evidence. While label propagation (LP) over inter-sample graphs captures propagation patterns, label scarcity limits its effectiveness. Recent methods inject LLM-generated pseudo-labels into LP to compensate, yet this naive integration introduces two issues: connecting all samples without selective distinction causes erroneous cross-class label propagation, and noisy pseudo-labels corrupt model parameters through a noise-amplification effect. To address these issues, we propose GCLP, a gradient-consistency-driven LP framework for multimodal fake news detection. We design Anchor-Driven Gradient Walk Graph Construction (ADGW), which leverages gradient consistency between high-confidence anchors and training nodes to build a class-aware propagation subgraph, eliminating erroneous label mismatch propagation. We further introduce a Gradient-Guided Denoising Masking mechanism (GGDM), which fuses per-parameter gradient signal-to-noise ratio with anchor-based consistency to selectively mask noisy updates, thereby protecting model parameters from label noise and mitigating the noise-amplification effect. Finally, Gradient Graph-based Label Propagation (GGLP) performs attention-weighted propagation over the augmented graph for the final prediction. Experiments on three benchmarks spanning two languages show that GCLP achieves state-of-the-art performance.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.