Advancements in artificial intelligence, and especially deep learning technology have given birth to a new era of multimedia forgery. Deepfake takes it to a whole new level. This deep learning based technology creates new images with features which have been acquired from a different set of images. The rapid evolution of Generative Adversarial networks (GANs) provides an available route to create deepfakes. They generate highly sophisticated and realistic images through deep learning and implement deepfake using image-to-image translation. We propose a novel, memory-efficient lightweight machine learning based deepfake detection method which is successfully deployed in the IoT platform. A detection API is proposed along with the detection method. To the best of the authors’ knowledge, this effort is the first ever for detecting highly sophisticated GAN generated deepfake images at the edge. The novelty of the work is achieving a considerable amount of accuracy with a short training time and inference at the edge device. The total time for sending the image to the edge, detecting and result display through the API is promising. Some discussion is also provided to improve accuracy and to reduce the inference time. A comparative study is also made by performing a three-fold textural analysis - computation of Shannon’s entropy, measurement of some of Haralick’s texture features (like contrast, dissimilarity, homogeneity, correlation) and study of the histograms of the generated images. Even when generated fake images look similar to the corresponding real images, the results present clear evidence that they differ significantly from the real images in entropy, contrast, dissimilarity, homogeneity, and correlation.