<p>Fake news on social media sites poses a significant threat to reliability, as it is multimodal and quickly spreads. In the existing works, the short news titles make it challenging to capture relevant semantic information. Current detection methods focus on examining the news image or headline alone, but limited work is on fusing multimodal features. Recent works in this area aim to develop Attentive Multi Fusion based Fake NewsNet (AMFNet) that concentrates on only one modality, like examining the news image or headline alone. This work applies the FakeNewsNet, Fakeedit, and Twitter MediaEval datasets, a benchmark dataset in the fake news detection field. The social media platforms involved in this work include Twitter, which is leveraged for its media content and social context data. A comprehensive multimodal technique is needed to design accurate fake news detection systems, incorporating user profiles and social context data. A hierarchical model using BERT and LSTM handles long text news bodies, a custom-designed CNN captures image tampering features, and an MLP captures social context features. The study also introduces an early feature fusion mechanism, leveraging scaled dot-product attention that synergistically combines visual, textual, and contextual features. Comparative analyses of early, intermediate, and hybrid fusion approaches demonstrate the effectiveness of multimodal integration over unimodal methods. The findings highlight the importance of feature fusion techniques in improving detection performance and reducing misclassifications.</p>