Deepfake technology has witness quick succession in previous years due to significant progress in deep learning and generative models, mainly Generative Adversarial Networks (GANs) and autoencoders. These technologies permit the foundation of extremely practical manufactured media, counting manipulated pictures and videos that are frequently hard to separate from genuine content. Whereas deepfakes offer valuable applications in areas such as entertainment, media generation,andvirtualsimulationsthey too present genuine dangers, counting deception spread, personality robbery, political control, and cybercrime. As a result, the improvement of consistent and computerized deepfake detection systems has suit a critical area of research in digital forensics and cybersecurity. This research proposes a productive and scalable deepfake detection system based on the EfficientNet architecture built-in with transfer learning techniques. EfficientNet is selected due to its ability to attain high accuracy with less restriction through its complex scaling method, which balances connection profundity, width, and determination. The utilize of transfer learning allows the model to influence pre-trained weights from large-scale datasets, enabling earlier convergence, reduced computational cost, and enhanced performance even with imperfect labeled data.

Keywords: Deepfake Detection, EfficientNet, Transfer Learning, CNN, Image Classification, Digital Forensics