The development of Generative Adversarial Networks (GANs) for generating realistic deepfake content through artificial intelligence brings a complex task to authenticate deepfake images. The spread of deepfakes leads to widespread distrust among people while simultaneously hurting both personal and community-based reputations through deceptive information distribution. This paper aims to develop an optimised deep learning model based on the EfficientNetV2-B1 architecture designed specifically for binary image classification of distinguishing real or fake. The proposed method has been implemented on the extensive 140k Real and Fake faces dataset from Kaggle. As other existing detection and classification approaches rely heavily on pre-trained models and a limited dataset, our model delivers compelling performance through a customised training methodology. As a result, the model was able to achieve 99.91% training and 98.76% testing accuracy coupled with the precision, recall and F1-score at 99.28%, 98.43%, 98% respectively. Furthermore, the model's performance is compared to the current techniques to show its reliability. The model's predictions are also interpreted using XAI visualisations, providing explainable insights into the areas of an image that contribute to its classification as either real or fake.