Deepfake images have been characterized by realistic manipulated images and videos, which have become one of the critical issues in the digital age which raises serious concerns about the integrity of the content. In this study, ELA-Net has been proposed, and this new type of framework which integrates error-level analysis (ELA) with deep learning and ensemble machine learning techniques for detecting deep fake. In this methodology, three stages have been considered for processing the deepfakes. In the first stage, preprocessing through multi-scale ELA where manipulated regions have been identified is followed by feature extraction by applying pre-trained Convolutional Neural Networks and classifying using ensemble learning methods such as Support Vector Machines (SVMs) and K-Nearest Neighbors (KNNs). The proposed model has been evaluated on the dataset where the 2041 real and fake images have been considered. Simulation results show that the proposed model has achieved a classification accuracy of 89.8%, precision of 91.34%, and an AUC-ROC score of 0.94, which is better than other pre-models such as ResNet18 + KNN and MesoInception4.