Abstract: Multimodal online rumors combining images and texts have become a prominent issue disrupting social order and affecting social stability due to their high deceptiveness and rapid spread. However, content-based detection methods often suffer from "negative transfer" and "seesaw" effects when processing multi-domain data, resulting in significant performance fluctuations across different vertical domains despite an increase in overall accuracy. To address this, an online Rumors Detection model fusing Hierarchical domain Experts and image-text Ambiguity Discrimination (RDHEAD） was proposed to enhance cross-domain adaptability while ensuring detection accuracy. First, an expert feature decoupling framework was designed to separate domain-general and domain-specific features by employing parallel sharing and unique expert networks combined with adversarial training and domain supervision. Second, an image-text ambiguity discrimination module based on cross-attention enhancement and Chinese CLIP fine-tuning was constructed to capture cross-modal conflicts. Finally, a two-stage gating network was utilized to dynamically fuse various features for rumor detection. Experimental results on the Weibo21 dataset showed that the macro F1 score of RDHEAD reached 92.8%, which is 0.5 percentage points higher than that of the CFPFND (multi-domain fake News Detection based on Cross-Feature Perception Fusion) model. Meanwhile, on the F1 Standard Deviation (F1-STD) metric evaluating inter-domain balance, RDHEAD is only 0.016, a 33% reduction compared to the state-of-the-art CCFMND (Cross-Correlation Feature and Memory-based Network for Detection) model. The experimental results verified that the proposed model effectively mitigated the "negative transfer" and "seesaw" problems in cross-domain rumor detection while improving detection accuracy.

Key words: rumor detection, multi-domain, multi-modal, domain experts, image-text ambiguity

摘要： 图文结合的多模态网络谣言凭借高迷惑性和快速传播能力，已成为扰乱社会秩序、影响社会稳定的突出问题。然而，基于谣言内容的检测方法在处理多领域数据时，常受制于“负迁移”与“跷跷板”效应，导致虽然总体判别精度上升但不同垂直领域的检测性能波动巨大。为此，提出一种融合层次领域专家与图文歧义判别的网络谣言检测模型(RDHEAD)，在保障检测精度的同时提升跨领域适应能力。首先，设计一个专家特征解构框架，通过并行的共享与独有专家网络分别进行学习，结合对抗训练与领域监督，实现领域通用与专属特征的分离。其次，设计一个基于交叉注意力增强与Chinese CLIP微调的图文歧义判别模块，以捕捉跨模态冲突。最后，通过两阶段门控网络动态融合各类特征，从而完成谣言检测。在Weibo21数据集上的实验结果表明，RDHEAD的宏平均F1值达到92.8%，比模型CFPFND(multi-domain fake News Detection based on Cross-Feature Perception Fusion)提升0.5个百分点；同时，在衡量领域均衡性的F1标准差(F1-STD)指标上，RDHEAD仅为0.016，相较于现有最优模型CCFMND(Cross-Correlation Feature and Memory-based Network for Detection)降低了33%。实验结果验证了所提模型在提高多领域谣言检测精度的同时，有效缓解了跨领域谣言检测的“负迁移”与“跷跷板”问题。

关键词: 谣言检测, 多领域, 多模态, 领域专家, 图文歧义