This summary outlines the experimental framework, performance evaluation, and robustness testing of a novel multimodal model designed to detect fake news. To ensure scientific reproducibility, the experiments were conducted on an NVIDIA RTX 3080 using the MATLAB Deep Learning Toolbox, with hyperparameters optimized via a Bayesian process to minimize cross-entropy loss. The model underwent rigorous 10-fold cross-validation on the GossipCop and PolitiFact datasets, consistently partitioning data into training, validation, and testing sets to prevent bias. By comparing various configurations—including character-level (CNG), word-level (W2V), and behavioral feature streams—the study demonstrated that integrating these modalities using a Multi-Head Self-Attention (MHSA) mechanism yields the most reliable results.
Performance measurements were analyzed across several metrics, including accuracy, precision, recall, F-measure, and the Matthews Correlation Coefficient (MCC). The results reveal that the proposed multimodal framework significantly outperforms existing state-of-the-art models. On the GossipCop dataset, the model achieved an accuracy of 99.17%, while maintaining stable performance on the more complex PolitiFact dataset at 96.85%. Confusion matrix analysis confirmed that the model exhibits exceptionally low type I and type II error rates, meaning it is highly adept at distinguishing between authentic and fabricated content across both domains.
A standout finding is the model’s efficiency compared to heavy architectures like RoBERTa-base. While transformer-based models are effective at parsing linguistic patterns, they often struggle with the non-textual metadata that characterizes modern misinformation. The proposed model, utilizing only 1.8 million parameters, outperformed RoBERTa-base in both accuracy and computational efficiency. This lightweight design allows for potential real-time application and deployment on edge devices, as evidenced by an average inference time of just 12 milliseconds per sample.
To address the limitations of “black box” deep learning, the study employed Explainable AI (XAI) by analyzing the weights of the MHSA layer. This analysis revealed that while the model balances semantic and statistical features for authentic news, it deliberately prioritizes behavioral and metadata signals (like repost-to-comment ratios) when flagging fake news. This indicates that the system is learning to identify the “behavioral fingerprints” of misinformation spreaders rather than relying solely on the text content, which can be easily manufactured to look legitimate.
Statistical rigor was applied to ensure the improvements were not merely coincidental. Using paired-sample t-tests, the researchers confirmed that the performance advantages of the MHSA mechanism were statistically significant (p < 0.05) compared to both baselines and existing literature. The tests clearly demonstrated that the intelligent fusion of character-level, word-level, and behavioral data is mathematically superior to any single-modality approach or simple concatenation of features, establishing the framework as a transformative tool for data integrity. Finally, the study tested the model’s robustness against real-world challenges such as noisy inputs, deliberate obfuscation, and domain shifts. By intentionally introducing typos and adversarial noise into the data, the researchers proved that the inclusion of character-level (CNG) and behavioral features makes the model remarkably resilient compared to traditional word-level models, which often fail when encountering out-of-vocabulary terms. Furthermore, the model’s ability to generalize across different topics—performing well on political data even when trained on celebrity gossip—suggests that it has successfully captured domain-agnostic patterns of information propagation, making it a powerful and versatile solution for combating the global spread of misinformation.