The exponential rise in deepfake technologies has introduced unprecedented threats to digital identity protection, social trust, and forensic authenticity verification. Deepfake generation methods, which once required sophisticated manual editing, can now be executed automatically using deep learning-based generative models. The increasing realism of manipulated facial imagery has rendered manual authentication unreliable, leading to renewed emphasis on automated detection systems. In this work, we propose a compact yet powerful convolutional neural network (CNN) architecture that balances high accuracy with low computational overhead. The designed framework targets resource constrained environments such as embedded systems, mobile platforms, and low-power devices. A curated dataset of 140,000 images equally divided between real and fake samples has been employed, involving a mixture of GAN based generation, face swapping techniques, and attribute modification pipelines. Multiple learning control strategies, including adaptive learning rate reduction and early stopping, are employed to stabilize convergence and reduce overfitting. Evaluation results demonstrate that the proposed approach achieves an accuracy of 96%, surpassing several conventional heavy models such as Xception and ResNet-50 under identical experimental conditions. The findings validate that efficient deepfake detection does not necessarily require computationally expensive networks, and that task-oriented architecture design can significantly improve inference efficiency while preserving model generality.