A New Approach for JPEG Steganalysis with a Cognitive Evolving Ensembler and Robust Feature Selection

In this paper, we develop a new robust feature selection scheme and an evolving ensemble classifier for stego content classification in a steganalysis framework. Steganalysis vs. steganography is a classical competition between two opposing research areas. Steganography focuses on hiding data within any media source such that the modified content becomes statistically indistinguishable from the original non-modified media. On the other hand, steganalysis focuses on detecting modified media that contains hidden data. Steganalysis includes two major steps, viz., feature extraction and binary classification of the original vs. modified images. The proposed Robust Feature Selection Method along with a Cognitive Evolving Ensemble classifier (RFSM-CEE) uses a Robust Feature Selection Genetic Algorithm (RFSGA) for identifying the robust features. A new measure called Sample Hardness (H) is used to calculate the Classifier Cost and select those training samples with higher sample hardness to train a set of basic classifiers with the robust features. RFSGA uses a specially tailored classifier cost C as the fitness function, which indicates the importance of each basic classifier for further ensembling. The proposed Cognitive Evolving Ensemble classifier (CEE) uses a growing/deleting strategy along with a voting scheme coupled with an Adaptive Ensemble Genetic Algorithm to define the set of basic classifiers for efficient ensembling. CEE uses simple voting rules to make a decision about each sample. Detailed performance evaluation of RFSM-CEE has been carried out by conducting experiments using J-UNIWARD and heuristic Bose-Chaudhuri-Hocquenghem steganography. The data used in these experiments are from BOSSbase and BOWS2 databases, along with Cartesian calibration JPEG Rich Models features. Experimental results clearly indicate major improvements in detection compared to the JPEG steganalysis ensemble classifier proposed by Kodovsky. In this paper a Robust Feature Selection Method along with a Cognitive Evolving Ensemble classifier (RFSM-CEE) focusing on searching for robust features in steganalysis data is presented along with a more accurate classifier to build efficient steganalysis.