The proliferation of DeepFake technologies has raised significant concerns across digital media, prompting the urgent need for reliable and efficient detection mechanisms. This study proposes a deep learning-based approach to DeepFake image detection using a custom-designed Convolutional Neural Network (CNN) implemented via Keras with TensorFlow backend. The model architecture features a hybrid of standard and separable convolutional layers, enhanced with Parametric ReLU activations, batch normalization, and dropout regularization to optimize learning and generalization. The model was trained on a curated dataset of facial images, categorized as either ‘real’ or ‘fake’, and evaluated using standard metrics across a 60-epoch training schedule. Experimental results demonstrate strong learning efficiency and robust generalization capabilities, achieving a final training accuracy of 92.45% and an exceptional validation accuracy of 99.98%, with correspondingly low loss values. These results were achieved with minimal training time, showcasing the model’s computational efficiency and effectiveness in binary image classification tasks. The findings validate the architectural and training design choices, and suggest that the proposed model offers a scalable and accurate solution for Deep-Fake detection. Future work may explore cross-dataset validation, real-time deployment, and integration of explainable AI for interpretability.