Fused Deep Representation of Traditional Features for Copy-Move Forgery Detection

Digital images are essential for displaying information. The rapid propagation of powerful editing tools increased the number of fake images intending to hide/duplicate sensitive information from the image. This complicates the task of discerning altered images from original ones. Developing effective techniques for detecting such forgeries becomes essential. One of the most widespread types of forgery is Copy-Move Forgery (CMF), which describes the action of copying and pasting a section of the original image into the designated area of that same image, that possesses the same properties as the original image, with easy and high-quality tampering. The forgery seems authentic since the target area has the same features as the original. Numerous methods to uncover CMF with a single feature exist. A drawback of such methods is their excessive execution time added to their low capability of detection accuracy. The presented paper introduces a copy-move image forgery detection model using deep learning, with two paths of feature extraction. The fused extracted features are given to the pre-trained MobileNet model to classify and detect the forgery in the images. The model was enhanced by replacing its classifier with a fine-tuned deep neural network classifier. The experimental results show that the proposed technique outperforms the state-of-the-art techniques by about 17.7%, 8%, 4.7%, and 2% in Recall, Precision, F1 score, and detection accuracy rate, respectively.