Indian Journal of Science and Technology

Abstract

Acquiring the best image features that best distinguishes a stego and clean image is a challenge in image steganalysis. Though higher order models acquire all these features, they pose problems due to computational complexity in terms of time and space. This demands optimization of the feature sets. Compared to the existing statistical feature optimization techniques, genetic algorithm based optimization techniques are evolving to be more promising. The existing deterministic methods of optimization have the limitation of converging into local minima as compared to the evolutionary methods which tend to converge to the global minima. Objectives: This paper intends to review the various genetic algorithm based feature optimization techniques applicable for image steganalysis of JPEG images and identify the best algorithm that converges to global minima. Method/Analysis: The methods analysed include the stochastic (metaheuristic) algorithms that make use of the random behaviour of plants and animals. The Antlion behaviour based optimization technique (ALO) has been implemented and analysed for JPEG stego images. The movement of ants are modelled as random walk and the traps built by antlions are assumed proportional to their fitness. The antlions shoot sand outwards to pull the ants inside the pits. This causes sliding down of the ants into the pits to the most minimum position. The coding of the optimization is implemented in Matlab with images taken from the standard BOSS database. Findings: The feature set after feature extraction has a dimension of 2000 × 48600 with 1000 cover and 1000 clean images. Considering these vectors as the initial positions of the ants in the Ant Lion Optimizer, for a payload of 0.5 in embedding logic the classification accuracies are studied. The convergence of this optimizer is proved according to the convergence curve for 300 iterations. After optimization, the reduced feature set is used to classify the image as cover or stego image. SVM, MLP and the fusion classifiers - Bayes, Decision template and Dempster Schafer are used. For low levels of embedding changes, the classification by MLP and Fusion schemes is good. For medium and high levels of embedding changes, the classification by Fusion schemes alone is good. It has been identified that the proposed steganalyser gives best results for Bayes fusion classification (69%) scheme when Antlion behaviour is used as optimizer. Applications/Improvements: This research has implemented a novel method of image feature optimization that improves steganalysis. The optimized feature set is 100 times less in dimension assuring reduced computational complexity in time and space. Improved version of this research may include a different selection mechanism or using a different optimization function. 

Keywords: Bio-Inspired Algorithms, Evolutionary Algorithm, Fusion Classifiers, Stochastic Optimization, Swarm Intelligence