Chronofactnet: A Memory-Augmented Multimodal Framework For Fake News Detection In Fakeddit Using Temporal Consistency Learning And Counterfactual Reasoning

Abstract

The volume and sheer speed of misinformation that can be spread on social media platforms has made it much more difficult to identify fake news, and misinformation is a multi-layered mix of text, false visual information and false metadata. Previous multimodal methods mostly focus on feature fusion and attention mechanisms, but often lack temporal inconsistencies and causal relationships between different modalities. To overcome these drawbacks, this research introduced a multimodal fake-news detection system that leverages temporal consistency learning and counterfactual reasoning (CFD), called ChronoFactNet. To circumvent these limitations, this work proposes a fake-news detection system based on memory-augmented multimodal framework, which trains a network using temporal consistency learning and counterfactual reasoning (CFD). The proposed architecture comprises six major components: (1) Textual Representation with DeBERTa, (2) Visual Feature Extraction using DINO-v2, (3) Cross-modal Semantic Alignment, (4) Temporal Memory Network, (5) Counterfactual Reasoning Module, and (6) Adaptive Truth-fusion Layer. After a thorough evaluation protocol (including comparative analysis, ablation studies, robustness and interpretability evaluation), experiments were run on the Fakeddit benchmark dataset. The proposed framework was able to achieve an accuracy of 95.84% and an ROC-AUC of 0.976 which is better than few state-of-the-art multimodal models. Moreover, the introduction of the Temporal Consistency Index, Counterfactual Stability Score and Memory Retrieval Effectiveness metrics illustrate the effectiveness of ChronoFactNet for capturing historical misinformation patterns and improve the reliability and robustness of the predictions, while also making them interpretable.