FuzConvSteganalysis: Steganalysis via fuzzy logic and convolutional neural network

Abstract

Emerging technologies based on the advancements in Deep Learning (DL) induced several alternative approaches to address intricate problems, such as analyzing images in the spatial to identify the location of the hidden content utilizing Convolutional Neural Networks (CNNs) as a backbone. Contemporarily, several CNN architectures have surfaced, elevating the accuracy of locating the concealed data in images. However, existing CNNs face challenges attributed to the heightened imperceptibility of the location of the secret data hidden with low payload capacities and less than optimal feature learning procedures. In this work, a steganalysis scheme named FuzConvSteganalysis proposes an innovative software tool that combines fuzzy logic and CNNs to locate the pixels holding the hidden information in the spatial domain images. FuzConvSteganalysis comprises three primary stages: the derivation of modification maps delineating alterations between the original image and the image containing concealed data, generating the correlation maps, and predicting the possible positions of hidden data. The maps resulting from the modification of the image serve as fuzzy inference system input and are subsequently fed into a CNN for classification. Through experimentation, FuzConvSteganalysis is assessed against four distinct adaptive data hiding approaches: WOW, HILL, S-UNIWARD, and HUGO-BD. Upon initial examination, the sensitivity for all four approaches exhibits a comparable upward trend, progressively enhancing with augmented payload capacity. The locating accuracy for the steganographically modified pixels attains a peak of 92.89% with WOW at a concealment rate of 0.5. This substantiates the superior efficacy of the FuzConvSteganalysis compared to the state-of-the-art algorithms.