The Deep Learning Shift In Steganalysis: A Modern Review

Along with the developments of adaptive steganography algorithms, it is increasingly difficult to detect data embedded in digital images. With the development of modern steganographic algorithms, such as substitution algorithm based on improvement models, frequency domain transformation and deep learning techniques, it is difficult to use traditional statistical features for revealing hidden patterns there. Conversely, there has been a boom of sophisticated analysis methods and in particular embedding large convolutional neural networks (CNNs) or residual learning models for multiple layers taking on different roles that exploit the subtle signals arising from steganography. This paper provides a new review on current trends in AI-based steganalysis, ranging from CNN schemes, through hybrids of deep learning with traditional classifiers, to domain-driven learning. It discusses current issues, such as differences between image sources, adversarial attacks, and expensive computational costs, as well as recent attempts to improve the generalization of embedding algorithms and emergent steganographic schemes’ dynamic adaptiveness.