<p>Clickbait refers to sensationalized and misleading headlines designed to attract user attention, often at the cost of informational credibility. Despite significant progress in clickbait detection for high-resource languages, low-resource languages such as Bangla remain severely underrepresented due to data scarcity and limited linguistic tools. Existing Bangla approaches further restrict their scope by focusing primarily on headlines, overlooking article-level context that is critical for robust detection. To bridge this gap, we propose a novel semi-supervised Generative Adversarial Network (SS-GAN) framework that advances Bangla clickbait detection by jointly modeling headline and full-article representations. Our approach introduces an attention-guided dual-representation fusion mechanism built on BanglaBERT, enabling effective exploitation of large-scale unlabeled data in resource-constrained settings. We further enhance semantic reliability by incorporating a cosine similarity–based consistency constraint between title and content representations within the discriminator. Comprehensive experiments demonstrate that the proposed framework achieves state-of-the-art performance, with an F1-score of 0.7595 and an accuracy of 0.8406. Beyond Bangla, this work establishes a scalable semi-supervised paradigm for clickbait detection in low-resource languages, contributing toward more trustworthy and resilient digital news platforms.</p>