Universal steganalysis of spatial content-independent and content-adaptive steganographic algorithms using normalized feature derived from empirical mode decomposed components

Abstract

Steganography is a practice of covert communication by hiding secret or payload information inside a cover source medium. Fraudulent entities may make use of it to communicate secretly and pose a great threat to law enforcement agencies. Universal steganalysis is a counter attack designed to identify steganography regardless of the steganographic process involved. As minute payloads that ride over covers, evade detection, steganalysts need to resort to high dimensional feature sets that extract residues in every possible way and a powerful classifier with higher computational power. In this paper, a novel universal steganalyzer driven by Empirical Mode Decomposition technique which is an adaptive and highly efficient method to analyze non-linear and non-stationary signals is proposed. Using this technique, the suspicious digital images are decomposed into intrinsic mode functions and residues to gradually remove the image contents to reveal the feeble artifacts caused by steganographic algorithms. Frobenius norm computed from these components further highlight the distortions. A low dimensional feature with a simple thresholding function assisting the steganalytic community in a blind and generic way can definitely be viewed as a boon to the domain. The proposed method is tested upon spatial content-adaptive steganographic algorithms (HUGO, WOW, S-UNIWARD) and content-independent least significant bit steganographic algorithms (LSBR, LSBM, LSBMR, LSBR2). The experimental results show that the proposed method is able to achieve improved accuracies when compared with Spatial Rich Model, Wu Net, Ye Net and Hybrid CNN. The execution time for the proposed method is also far less compared to state-of-the-art steganalyzers.