Discriminative Algorithms for Large-Scale Image Steganalysis and Their Limitations

Digital image steganography is the art and science of hidden information. Currently, steganographic (stego) algorithms are rapidly evolving and reducing their artifacts. There-fore, detecting of altered cover images, i.e. steganalysis, is more challenging. Modern steganalysis is based on machine learning techniques that make decisions based on training information. We have found that those methods do not generally work under real-world conditions, where the training and testing image datasets are numerous. Moreover, we will show that the current methods produce unpredictable results. That is, if the methods work well under a dataset, they are not necessary work well on a different dataset. In this thesis, we show that steganalysis based on discriminative approaches cannot be in-dependently used to detect steganographic images, and we provide their limitations. Thus, we should look for alternative approaches. Additionally, we propose a generative model approach to steganalysis for detecting steganographic images among a large number of im-ages, acknowledging that most images are intact. The system consists of a series of intrinsic image formations filters (IIFFs), where filters are designed to detect non-steganographic images based on real-world constraints and intrinsic features of steganographic methods. Our approach can be used as a basis for building a robust and reliable steganalytic system.