The spread of fake news or hoaxes remains a significant issue in today’s digital world. Detecting fake news becomes even more challenging when dealing with limited data. This study looks at how to use data augmentation techniques such as Synonym Replacement (SR) and Random Insertion (RI) to improve the performance of fake news classification models. Accuracy and F1 Score are used to evaluate the effectiveness of these techniques at three different augmentation levels: low (0.25), medium (0.5), and high (1). The results of this study show that SR and RI significantly enhance model performance than the original dataset. SR outperforms RI due to its ability to preserve semantic meaning, especially at higher augmentation levels. On the other hand, RI makes some changes that generalize better, especially at lower augmentation levels. The ISOT Dataset (Dataset 1) has 23,502 fake and 21,417 real news articles for binary classification. The WELFake Dataset (Dataset 2) contains 72,134 articles (35,028 real, 37,106 fake), merging multiple sources to improve diversity and reduce overfitting. Experimental results show that the Accuracy score increased from 0.81 to 0.93 and the F1 Score improved from 0.82 to 0.93 on Dataset 1, and for Dataset 2, the Accuracy increased from 0.81 to 0.93, and the F1 Score improved from 0.80 to 0.93. However, the techniques have limitations, as they have not yet been tested in real-world applications, where domain-specific challenges and scalability could impact their performance. Overall, the results show that data augmentation gives positive improvements over the fake news classification task by solving issues related to imbalanced datasets and enhancing the model's ability to generalize effectively.