Universal Steganalysis Through Meta Heuristic Optimization and Deep Learning Techniques

Abstract

Steganography techniques of hiding information in digital images have been found to newlinebe predominantly used for various malicious acts like cybercrime, and terrorism. Therefore, newlinesteganalysis methods that can counter-attack steganography have gained importance. These newlinemethods rely upon the extraction of thousands of higher-order statistical features to train machine newlinelearning classifiers which aid in the detection of an image as a cover (clean image) or a stego (image newlinewith hidden data). This study focuses on improving steganalysis to detect stego images created by newlineadvanced steganography methods through the emerging techniques of machine learning and deep newlinelearning. newlineInitially, a comparative analysis of different classifiers is done on the Spatial Rich Model (SRM) newlinefeature set of 106 feature submodels extracted from the dataset of cover and stego images. The newlinecombined features of 106 feature submodels is 34671. This analysis helps to conclude that the newlinecomputationally fast ensemble of the Fisher Linear Discriminant (FLD) classifier performs better than newlinethe other classifiers used. However, the memory requirement and feature extraction time for a high- newlinedimensional dataset is quite high. Further, high-dimensional datasets may contain many redundant newlineand irrelevant features, leading to the suboptimal performance of the various classifiers. Hence, to newlineaddress these challenges, effective feature selection techniques are required to reduce high- newlinedimensional features. Two different insights of feature selection have been utilized in this work. A newlinemodified meta-heuristic method based on Particle Swarm Optimization (PSO) has been applied for newlinethe selection of optimal feature submodels out of 106 SRM submodels. Next, a hybrid meta-heuristic newlinemethod based on Harris Hawks Optimizer (HHO) has been developed and applied to the combined newline34671 SRM features. Both methods have resulted in a significant reduction of feature space and newlinecomputational time using the ensemble classifier while maintaining comparable detection accuracies newlinewith the complet