Steganalysis with mismatched covers: do simple classifiers help?

Authors:
Ivans Lubenko;Andrew D. Ker
Affiliations:
University of Oxford, Oxford, United Kingdom;University of Oxford, Oxford, United Kingdom
Venue:
Proceedings of the on Multimedia and security
Year:
2012

Citing 10
Cited 0

Large Margin Classification Using the Perceptron Algorithm

Machine Learning - The Eleventh Annual Conference on computational Learning Theory
Restricted Bayes Optimal Classifiers

Proceedings of the Seventeenth National Conference on Artificial Intelligence and Twelfth Conference on Innovative Applications of Artificial Intelligence
Statistically undetectable jpeg steganography: dead ends challenges, and opportunities

Proceedings of the 9th workshop on Multimedia & security
The Unreasonable Effectiveness of Data

IEEE Intelligent Systems
Kernel methods in steganalysis

Kernel methods in steganalysis
Steganalysis by subtractive pixel adjacency matrix

Proceedings of the 11th ACM workshop on Multimedia and security
LIBSVM: A library for support vector machines

ACM Transactions on Intelligent Systems and Technology (TIST)
"Break our steganographic system": the ins and outs of organizing BOSS

IH'11 Proceedings of the 13th international conference on Information hiding
A new methodology in steganalysis: breaking highly undetectable steganograpy (HUGO)

IH'11 Proceedings of the 13th international conference on Information hiding
Steganalysis of content-adaptive steganography in spatial domain

IH'11 Proceedings of the 13th international conference on Information hiding

Quantified Score

Hi-index	0.00

Visualization

Abstract

Modern steganalysis can be incredibly sensitive and accurate, but only in an artificial setting in which the training data is from the exact same image source as the data being tested. When faced with mismatched training data, performance degrades, often substantially. Some recent publications have advocated the use of very simple classifiers for steganalysis. It is folklore from the machine learning literature that simple classifiers may be more robust to variations between training and testing data, and this paper investigates whether this is true for steganalysis also. Simple classifiers are compared with the classic complex Kernel Support Vector Machine, faced with training data which is mismatched with the testing data. A real-world source of images is used.