Social media platforms are sometimes held responsible for the spread of fake news all over the Internet, and thus there comes the urge and need for the early detection of these types of fake news in order to increase the quality of information and data. In this paper, we propose three deep learning-based fusion models that are used to detect fake news. The fusion models include a combination of CNN, BiLSTM, attention mechanisms, and the HuggingFace BERT transformer models. We have experimented with the CNN-BiLSTM-Attention-based fusion models in two flavors. The first one included the encoding of individual characters and the second one employed word-tokenization method. The performance of the HuggingFace BERT model outperformed the other two models by obtaining the highest accuracy with a 0.99 F1-score for both fake and real news detection. The word-level fusion prototype of CNN-BiLSTM performed moderately well with a F1-score of 0.95. However, the character-level CNN-BiLSTM model struggled a bit with a F1-score of 0.90. These findings highlight the overall superiority of the transformer-based model for fake news detection.