The present work explores the effectiveness of data augmentation techniques in improving deep learning models for image classification. Deep learning models suffer from overfitting when labelled data is limited, limiting their ability to generalize effectively. Conventional techniques such as regularization and dropout can help to some extent, but data augmentation provides a novel approach to addressing this issue. The paper utilizes a small set of images from the Tiny ImageNet-200 dataset, which has only 500 images per class. Initially, traditional data augmentation techniques such as rotations, zooming, and flipping are used to vary the dataset and reduce overfitting. Following that, the study goes into Generative Adversarial Networks (GANs) for data augmentation, employing a Deep Convolutional GAN (DCGAN) to generate synthetic images that closely resemble real data. The experimental results show that data augmentation has a considerable impact on model performance. The study describes three experiments: a baseline model with no data augmentation, a model trained with traditional augmentation, and a model with both traditional and GAN-based augmentation.