Fake news detection in underrepresented languages like Bengali faces significant challenges due to limited annotated data and severe class imbalance. This paper presents a novel hybrid approach that integrates extractive and abstractive summarization to generate diverse fake news samples, enriching the training dataset. (YAKE!) is used for extractive summarization to identify key terms, while (mT5) performs abstractive summarization to rephrase content while maintaining its meaning. The newly generated summaries are added to the dataset, making it more balanced and improving model generalization. Unlike SMOTE or ADASYN, which generate synthetic numerical samples, hybrid summarization enhances textual diversity while preserving semantic integrity. For feature extraction, we employ BanglaBERT and mBERT, capturing both language-specific and multilingual contextual representations. Their embeddings are concatenated to enrich feature representation, improving differentiation between fake and authentic news. The classification model is trained using concatenated Embeddings, enabling it to learn complex text dependencies. Additionally, class-weighted training ensures the model adequately focuses on the underrepresented fake news class. Evaluation on BanFakeNews and BanFakeNews 2.0 datasets demonstrates significant improvements over baseline models. Our approach enhances dataset quality, improves feature extraction, and strengthens classification accuracy. The findings highlight the effectiveness of hybrid summarization and Transformer-Based Embeddings in addressing fake news detection challenges in resource-constrained languages.