Real-time estimation for the parameters of Gaussian filtering via deep learning

Abstract

Driven by the development of digital technology, manipulation towards digital images becomes simpler than ever before in recent years. Many smartphone applications bring the convenience for ordinary people to edit images in real-time without any professional skills. The digital forensics is an important research field in information security against the situation. In image forensics, it is necessary to validate all possible manipulation during the forming history of given images. Thus, many image forensics researchers focus on detecting certain manipulations to protect the integrity of images such as verifying Gaussian filtering. However, these works tend to make binary classification that if the image is processed by certain manipulation or not. The classification of same manipulation based on parameters are ignored. Here, we propose a method to estimate the parameters of Gaussian filtering to process images based on convolutional neural networks (CNN). Besides, in the modern world, it is also extremely important to enable the simulation in real-time to process with the given data immediately. The proposed method can also validate the given image in a quite short time. Our experiments show that the proposed method can provide excellent real-time performance in estimating the window size and standard deviation of Gaussian filterings. The well-trained model can satisfy us with not only the estimation accuracy, but also the validation time simultaneously.