An Ensemble Deep Learning Models Based on Metadata for Measuring Arabic Fake News Uncertainty

Fake news, misinformation, and disinformation are significant cybersecurity threats worldwide, as sharing misleading or biased information can have detrimental consequences, leading individuals to act in harmful ways. It is crucial to provide fact-checking tools capable of determining the authenticity of news content, images, and videos. Despite the increasing penetration of fake news in Arabic countries, fake news detection in this context has received limited attention. This work focuses on developing an efficient approach to accurately identify fake news while also measuring the uncertainty of its predictions, a crucial aspect that has been overlooked in the field. We propose a powerful deep ensemble model that incorporates various multi-input merging techniques and different architectures. Our proposed architecture comprises a deep ensemble of three distinct models, each designed to handle specific types of inputs. The ensemble consists of Bi-Long Short Term Memory (Bi-LSTM) model utilizing embedded inputs, a Deep Neural Network (DNN) model incorporating only metadata, and a DNN model employing the Term Frequency-Inverse Document Frequency (TF-IDF) matrix. Additionally, we compute uncertainty of this detection. We further utilize a Bayesian Neural Network (BNN) architecture specifically for metadata to independently measure uncertainty and to give a qualitative evaluation of our uncertainty measure. Experimental results on a challenging dataset demonstrate that our proposed ensemble model outperforms previous works in terms of per-formance. Furthermore, by leveraging the BNN, we showcase how our ensemble architecture surpasses the standalone BNN in uncertainty accuracy. These results not only establish the validity of our proposed ensemble architecture for fake news detection, but also provide valuable information about the uncertainty associated with the model's predictions. By empowering internet users to become more critical consumers of news, these insights hold immense potential in mitigating the spread and impact of fake news.