An improved dynamic discrete firefly algorithm for blind image steganalysis

This paper presents an improved firefly algorithm (DyFA) for feature selection that improves the convergence rate and reduces computational complexity through dynamic adaptation in blind image steganalysis. The alpha and gamma parameters of the Firefly algorithm are made to vary dynamically with each generation for faster convergence. If firefly algorithm's performance does not improve for certain numbers of iterations then the particles with the worst fitness function values are replaced with new particles in the search space and particle dimensions are reduced by eliminating redundant features. This approach is effective in reducing computational complexity and improving detection capability of the classifier. To further reduce the computational complexity a hybrid DyFA is designed by ensemble of a filter approach (t test + regression) and wrapper approach (DyFA) incrementally. In this study, support vector machine classifier with radial basis function kernel and ten fold cross validation is used to evaluate the effectiveness of the proposed Firefly algorithm. DyFA is compared with well-known wrapper feature selection algorithms. Experimental results are performed on datasets constructed from four steganography algorithms nsF5, Perturbed Quantization, Outguess and Steghide with subtractive pixel adjacency matrix (SPAM) feature vector from spatial domain and Cartesian Calibrated features extracted by PevnA 1/2 feature vector from transform domain. Experimental results demonstrate that DyFA reduces computation time and improves classification accuracy as compared to other feature selection algorithms. Hybrid DyFA shows an improvement in classification accuracy and in eliminating redundant features in more than 85 % of cases with respect to hybrid GLBPSO.