Forgery detection in digital images using the hybrid deep learning method

Today, images are used as powerful communication tools and sources of information. In certain applications, such as medicine, justice, and forensics, images serve as evidence. Therefore, the validity of an image is crucial. With the spread and availability of image editing tools, people can easily manipulate images to their advantage. They follow political, cultural, economic, and social issues by adding or removing elements from images, often distributing misinformation. Consequently, forgery detection is one of the most important and challenging topics in the field of computer vision. This research aims to identify forgery and healthy images and pixels using a hybrid deep learning network. In the proposed method, three pre-trained networks—VGG16, MobileNet, and EfficientNetB0—are employed in three different branches. To detect forgery at both the image and pixel levels, the output feature maps from these branches are merged in a concatenate layer. Subsequently, a global average pooling layer and a scoring layer are used to identify forgery and healthy images. Additionally, feature maps combined from the three branches are utilized to create a heat map image for forgery detection. Notably, pixel forgery detection is performed solely using the heat map image generated from the combined network, without relying on ground truth images that specify the forgery area during training. The proposed method is evaluated on the well-known CoMoFod dataset, demonstrating satisfactory performance against forgery images with various geometric transformations and post-processing operations