Deepfake Detection Technology: A Review Of Lightweight Transfer-Learning Detectors, Artefact Cues And Generalization To Unseen Manipulations

Deepfake technology uses deep learning to create or manipulate images and videos so that a person appears to say or do something that never happened. Although it has legitimate uses, it can spread misinformation, enable fraud and damage reputations. This paper reviews deepfake detection technology with a focus on lightweight, deployable detectors and the generalization problem. It examines how manipulated faces are produced, the benchmark datasets used for training and testing, and detectors based on transfer learning, in which convolutional networks such as MobileNetV2 and EfficientNet pretrained on ImageNet are fine-tuned to classify face images or video frames as real or fake. It reviews artefact cues that detectors exploit, including blending boundaries, frequency-domain traces of up-sampling and fine-grained texture differences captured by attention, and the finding that detectors trained on one generator can transfer to others when trained with suitable augmentation. Detection of images from diffusion models is also discussed. A five-stage architecture comprising media input, face extraction and preprocessing, transfer-learning feature extraction, classification and a user interface built with a lightweight web framework is proposed, together with research gaps, objectives, a comparison of detector backbones and challenges such as compression, dataset bias and adversarial evasion. The review concludes that predictions should be treated as indications rather than proof and that robust generalization remains the main open problem.