With sudden proliferation in evolving Deepfake technology to manipulate media content, identity theft etc., several modern deep learning based solutions are being continuously developed for Deepfake detection, distinguishing the manipulated media. Therefore, the work presented in this paper proposes a novel and real-time MobileNetV3 based Deepfake detection model. With the fine-tuning of MobileNetV3 using additional fully connected layers and self-attention mechanism placed optimally, the proposed model proved effective in concentrating on crucial image regions over relatively lesser important features. The integration of self-attention layers had a direct impact on the model’s capacity to identify the artifacts manipulated during the generation of Deepfakes, thereby, alleviating the detection efficiency. The work also implemented and trained five widely used and recent convolutional models including XceptionNet, InceptionV3, EfficientNetB0, MobileNetV3, and InceptionResNet for best baseline model selection and comparisons against the proposed approach. The proposed model outperformed all other state-of-art models by a margin of 3% with an F-score = 98.46%. The proposed model also has minimalistic memory and time requirements supporting the real-time applicability of the model even on mobile devices.