A federated deep learning framework with distributed hybrid character-level and attention mechanisms for scalable and...

Fake news detection is an essential task for media and news organizations to maintain the trust and reliability of the published content. Due to the rapid growth of online users and the spread of misinformation through malicious sources, the fake news circulates quickly across digital platforms. Hence, developing an accurate and efficient fake news detection model is crucial for social welfare. Although many studies have been conducted, achieving high accuracy remains challenging task, especially when dealing with rare and infrequent words. Traditional embedding models such as Word2Vec and FastText perform only word-level vectorization, while transformer-based models like BERT handle subword token more effectively. However, selecting the most suitable subword pieces for vector representation raises challenges. The research introduces a character-based embedding approach to generate fine-grained vector representations by analysing each character. A hybrid CharBERT-Optimized CNN and attention-based stacked Bi-LSTM with cost-sensitive learning (COASBC) model is proposed to classify fake news. In addition, a federated learning (FL) framework is integrated to enable distributed training across multiple news sources for improving generalization of model without sharing raw data. It reduces centralized computational cost, enhances data security and scalability of the model. Furthermore, statistical analysis and complexity analysis are conducted to evaluate the efficiency and computational feasibility of the proposed model. The generalization ability of the model is also analysed using Truth Seeker dataset 2023 that contains short-text tweets. Experimental evaluation using the LIAR, Fake and Real News datasets, FakeNewsNet, and WELFake Dataset demonstrates that the proposed Federated COASBC model achieves superior accuracy, precision, and reliability compared to existing state-of-the-art fake news detection methods and it maintains high computational efficiency.