An Attentive Fusion Framework For Multi-modal Misinformation Detection

With the growing influence of social media, misinformation has become increasingly prevalent in recent years. Misinformation on social media often appears in multimodal form (i.e., image and text) to increase its credibility and reach. While there has been progress in multimodal fake news detection, most methods struggle to capture the complex interactions between text and images, leading to inaccurate results. This work proposes an attentive fusion framework for detecting multimodal fake news. Our approach employs ResNet50 to extract visual features and BERT for textual representations, followed by an attention layer to integrate the modality-specific features for fake news detection. We evaluated our method on two benchmark datasets-Gossipcop and Politifact- and achieved significant improvement with an accuracy of 88.55% and 88.46%, respectively. This outcome reflects around 2.9% and 4.6% gains compared to the state-of-the-art models.