Details

Autor(en) / Beteiligte

• Uppada, Santosh Kumar
• Ashwin, B. S.
• Sivaselvan, B.

Titel

A novel evolutionary approach-based multimodal model to detect fake news in OSNs using text and metadata

Ist Teil von

• The Journal of supercomputing, 2024, Vol.80 (2), p.1522-1553

Ort / Verlag

New York: Springer US

Erscheinungsjahr

2024

Quelle

SpringerLink

Beschreibungen/Notizen

• Social networks are the primary source of information sharing, and with the abundant availability of the internet, more and more information is being dumped into social networks from various sources. There is no proper credibility check, so social networks have become the main source for manipulated or inclined news. As information comes in multiple forms, there is a need for multimodal models to detect Fake News. Many features being specific to the domain may mislead the process of classifying news. Obtaining specific features that could retain the framework’s capability becomes the point of interest. The paper proposes a novel method first to obtain features from a specific attribute and use a mechanism to derive optimal features that helps in attaining the ability to detect Fake News across domains. Genetic algorithms, termed nature-driven, aim at exploring the best optimal features that enhance the detection capability of the framework. Therefore, the paper proposes a model that works on textual features, metadata, and author embedding data and uses genetic algorithms to extract optimal features. These multimodal are combined using fusion methods, and finally, 2-Layer MLP is used to detect the Fake posts. Fakeddit, a multimodal dataset, is used for analysis. Text features are generated from the Title parameter, which is used as clickbait to attract attention. Genetic algorithms like two-phase gray wolf (TMGWO) and improved sine cosine algorithms (ISCA) are used to extract the best optimal features from the dataset. The proposed model has achieved an accuracy of 91.05% when ISCA is used to extract optimal features, BERT is used to analyze textual features, and metadata and author embedding are combined using Concatenate. The model has shown that extracting optimal features has enhanced the overall accuracy of the multimodal model. The proposed model is observed to outperform other state-of-the-art models working on similar data. To the best of our knowledge, the proposed framework is prominent in obtaining specific features that could help to classify news or articles as real or fake using the blend of Genetic-based and fusion-based models.