MDCV : Multimodal Deceptive Clickbait Video Detection Using Transformer‐Based Model in Bangla

The proliferation of internet usage has made people increasingly dependent on it. Some fraudulent content makers make the content in an intriguing way that engenders users' curiosity. However, after clicking, they are deceived, which wastes users' time and degrades the credibility of the content. The majority of the existing clickbait detectors are title‐based and article‐based, but billions of people use video platforms daily. A few studies have been conducted on low‐resource languages, such as Bangla, due to the scarcity of resources. Therefore, to address this issue, we proposed a Bangla multimodal clickbait video detector that integrates video title and thumbnail and attained 94% precision. The model is trained and tested on a curated Bangla multimodal clickbait video dataset consisting of 253 570 samples. In this research, we used BanglaBERT followed by a BiLSTM model to maintain the rich contextual semantics for title encoding and ResNet50 with CBAM to capture the video thumbnail's prominent feature. Furthermore, we use cross‐modal attention to align the title and thumbnail data and an ablation study to ensure the contribution of each module. The results underscore the efficacy of each module and validate the effectiveness of our approach in detecting clickbait content across modalities.