Fake news detection in low-resourced languages has become a critical challenge in our information-driven society. Contextualized word embeddings can improve the accuracy of fake news detection in low-resource languages compared to traditional word embedding methods. However, seizing the complex relationships among words in news articles is challenging when using a specific contextualized word embedding strategy. To address this issue, we propose a computationally efficient heterogeneous ensemble-based Dravidian fake news detection framework that utilizes gradient accumulation to enhance base-level and meta-level classifiers. This framework employs three optimized BiLSTM base-level classifiers, each paired with different contextualized word embeddings: multilingual BERT, XLM-RoBERTa, and MuRIL. The predictions from these base classifiers are integrated using an attention mechanism and fed into an optimized LSTM meta-classifier for final classification, ensuring a comprehensive and context-aware decision-making process. On the Dravidian_Fake dataset (Telugu, Kannada, Tamil, and Malayalam), our framework achieves significantly higher accuracy and F1 scores compared to baselines, demonstrating its effectiveness in capturing complex word relations and robustly detecting fake news in low-resource settings.