Nowadays counterfeiting of banknotes, driving licenses, governmental and non-governmental documents is a major challenge to process business activities and identifying the person's abilities, which make forgery detection the hot research area. Ethiopian Driving License is a permit issued in accordance with Ethiopian proclamation number 1074/2018 to drive a motor vehicle. Driving with fake Driving License is a criminal activity and a prevailing factor for traffic accidents, which results in injury or death of human life and destruction of properties. Previously many researchers proposed a fingerprint-based system to verify driving license. Applying those technologies to developing countries like Ethiopia is difficult, since it requires high initial investment. This study aims to design a model that automatically detects fake Driving Licenses. For our study we are collecting fake Driving License from West Gondar Road Transport Office and Bahir Dar Traffic Police Offices, and genuine licenses are collected from different Driver’s having genuine license around Bahir Dar town and from Bahirdar Road Transport Office. The images are captured using a smartphone camera. After the images are prepared, we used Convolutional Neural Network (CNN) to extract deep features and Oriented FAST and Rotated BRIEF (ORB) to extract local features. In order to extract the best local feature extractor, we are performed a series of experiments on SIFT, SURF, and ORB, and we are achieved an accuracy of 86.27%, 86.23%, and 88.23% respectively. A series of experiments is also performed in order to select CNN parameters. After that we are merging the ORB local feature vector with the CNN feature vector in order to get the more discriminative feature of fake driving license, then SVM is used in order to classify the merged feature vector. When we used end-to-end CNN, ORB-CNN feature vector with SVM classifier we achieved an accuracy of 90% and 92% respectively. So, merging local features with CNN deep features results in good performance rather than using them individually.