Exploring the Power of Multimodal Fusion Strategies for Fake News Detection: A Comparative Analysis of Basic and Advanced Approaches

The widespread use of social media has changed how people access news and information. At the same time, it has made it easier for false information to spread quickly. There is an increasing concern about multimodal fake news, which uses both text and images, often edited or meant to be misleading, to deceive people. These fake stories are becoming more sophisticated and harder to detect, making it important to develop detection models that focus on how text and visuals work together, not just on each modal separately. This paper presents a comparative study between two intermediate fusion methods for detecting fake news that combine text and images—ConcatFusion, which uses basic feature concatenation, and AlignFusion, which builds on this by applying cross-attention to better capture the connection between the two modalities. The proposed AlignFusion method improves performance by learning fine-grained correlations between modalities, emphasizing the importance of explicitly modeling text-image interactions. Both fusion methods are tested using the FakeEdit dataset. To understand how different image models affect performance, three advanced visual encoders—EfficientNetB0, M obileNetV3, and ConvNeXt—are each paired with DistilBERT for text. This setup helps extract useful features from both images and text, making it easier to uncover how they work together in detecting fake news. The results show that AlignFusion consistently performs better than ConcatFusion. The best results are achieved by combining ConvNeXt and DistilBERT using AlignFusion, reaching an F1 score of 90.4%. These findings highlight the value of cross-attention-enhanced fusion in capturing complex relationships between text and visuals, ultimately leading to more reliable fake news detection.