A real-time image forensics scheme based on multi-domain learning

In recent years, researchers have attempted to explore methods for real-time image forgery detection. Many approaches were developed to detect a certain number of image modification methods. There are many limitations in practical application. In this paper, a multi-domain learning convolutional neural network (MDL-CNN) is proposed to overcome this limitation. We extract the periodicity property from the original and modified image. Features of modified image extracted from different datasets are then fed into the neural network in training process. Since the proposed MDL-CNN is trained by different types of tempering datasets, our method can distinguish many types of image modifications. To decrease the computation of proposed scheme, 1 x 1 kernel convolution layer is used in the second convolutional layer of each network. Furthermore, a multi-domain loss function is developed to enhance the recognition ability of in-depth learning features. Experimental evaluation results show that MDL-CNN method can significantly improve the forensic performance.