Deep learning approaches to universal and practical steganalysis

Abstract

Steganography is the process of hiding data inside of files while steganalysis is the process of detecting the presence of hidden data inside of files. As a concealment system, steganography is effective at safeguarding the privacy and security of information. Due to its effectiveness as a concealment system, bad actors have increasingly begun using steganography to transmit exploits or other malicious information. Steganography thus poses a significant security risk, demanding serious attention and emphasizing a need for universal and practical steganalysis models that can defend against steganography-based attack vectors. In this thesis, we provide a comprehensive review of steganography-enabled exploits and design a robust framework for universal and practical deep-learning steganalysis. As part of our framework, we provide new and practical steganalysis architectures, propose several data augmentation techniques which includes a novel adversarial-attack system, and develop a python library, StegBench, to enable dynamic and robust steganalysis evaluation. Altogether, our framework enables the development of practical and universal steganalysis models that can be used robustly in security applications to neutralize steganography-based threat models.