Share Email Print
cover

Proceedings Paper

Application of genetic algorithm to steganalysis
Author(s): Timothy Knapik; Ephraim Lo; John A. Marsh
Format Member Price Non-Member Price
PDF $14.40 $18.00

Paper Abstract

We present a novel application of genetic algorithm (GA) to optimal feature set selection in supervised learning using support vector machine (SVM) for steganalysis. Steganalysis attempts to determine whether a cover object (in our case an image file) contains hidden information. This is a bivariate classification problem: the image either does or does not contain hidden data. Our SVM classifier uses a training set of images with known classification to "learn" how to classify images with unknown classification. The SVM uses a feature set, essentially a set of statistical quantities extracted from the image. The performance of the SVM classifier is heavily dependent on the feature set used. Too many features not only increase computation time but decrease performance, and too few features do not provide enough information for accurate classification. Our steganalysis technique uses entropic features that yield up to 240 features per image. The selection of an optimum feature set is a problem that lends itself well to genetic algorithm optimization. We describe this technique in detail and present a "GA optimized" feature set of 48 features that, for our application, optimizes the tradeoff between computation time and classification accuracy.

Paper Details

Date Published: 22 May 2006
PDF: 7 pages
Proc. SPIE 6228, Modeling and Simulation for Military Applications, 62280X (22 May 2006); doi: 10.1117/12.669088
Show Author Affiliations
Timothy Knapik, SI International (United States)
Ephraim Lo, SI International (United States)
John A. Marsh, SI International (United States)


Published in SPIE Proceedings Vol. 6228:
Modeling and Simulation for Military Applications
Kevin Schum; Alex F. Sisti, Editor(s)

© SPIE. Terms of Use
Back to Top