IMFND-AFR: an incomplete-modality fake news detection framework based on active feature reconstruction

Abstract

Multi-domain multimodal fake news detection has attracted increasing research attention because misinformation on social media often involves both textual and visual content across heterogeneous topical domains. However, most existing methods assume that textual and visual modalities are simultaneously available, which limits their applicability in realistic scenarios where images may be missing because of privacy restrictions, crawling failures, platform deletion, or transmission errors. Passive imputation strategies, such as zero-padding, may introduce feature distribution mismatch and fail to recover domain-relevant visual information. To address this problem, we propose IMFND-AFR, an incomplete-modality fake news detection framework based on active feature reconstruction. The framework contains three main components. First, an orthogonality-constrained dual-stream expert module disentangles cross-domain general features and domain-specific features. Second, a domain-guided cross-modal reconstruction module uses textual semantics and domain embeddings as conditional priors to reconstruct missing visual general features, visual domain-specific features, and vision-language aligned visual features. Third, a complementary specificity-injection fusion strategy integrates domain-specific cues into general representations and dynamically combines real or reconstructed visual features under different modality conditions. Experiments on the Weibo and Weibo-21 datasets show that IMFND-AFR achieves complete-modality accuracies of 0.941 and 0.947, respectively, and maintains stable performance as the visual missing rate increases up to 80%. Compared with the zero-padding baseline, IMFND-AFR consistently achieves higher accuracy under visual-missing settings. Additional comparisons with mean imputation, prototype imputation, and modality dropout further show that active reconstruction achieves the best accuracy at all visual missing rates on Weibo-21 and the highest average accuracy across missing rates on Weibo. These results indicate that active feature reconstruction can improve the robustness of multi-domain multimodal fake news detection when visual information is incomplete, while its marginal benefit depends on the informativeness of visual cues in the dataset.