Statistical steganalysis detector model for 8-bit depth images

Other Title(s)

نموذج إحصائي لكشف الإخفاء في الصور ذات عمق 8-بت

Dissertant

al-Tai, Zayd Hadi

Thesis advisor

al-Jarrah, Muzaffar

Comitee Members

Abu Arqub, Abd al-Rahman
al-Najdawi, Nijad

University

Middle East University

Faculty

Faculty of Information Technology

Department

Computer Science Department

University Country

Jordan

Degree

Master

Degree Date

2017

English Abstract

This thesis aims to develop a statistical model for steganalysis to enhance the detection of the existence of hidden data inside 8-bit depth gray-scale BMP images.

The proposed model is based on enhancing the image texture features through analyzing both full-bytes and parts of bytes of an image.

It is known that most steganography techniques embed the bits of a secret message within the right half of a cover image’s bytes, the least significant half of a byte, to avoid obvious visual distortion.

Therefore, the focus of the steganalysis process in this work is the right half part of each byte of an image under investigation.

The selected feature set is based on the gray level co-occurrence model, including contrast, homogeneity, correlation, and energy.

Additional features include: entropy, coefficient of variation of the image’s right half-bytes, correlation coefficient between left and right half-bytes, and the average of difference between the intensity of right half -bytes in successive pixels.

The work involved implementation of the proposed model in MATLAB, which consisted of modules for feature extraction, training and testing using the two-category discriminant analysis classifier, and the batch classification of a set of test images.

Testing of the detection accuracy of the proposed model was carried out in three stages.

First, a dataset of 180 mixed-source images were analyzed using 3-fold cross validation.

In the second testing stage, the 3-fold cross validation was applied using a dataset of 1500 images from a single public dataset.

The third stage realized a large-scale field test of the proposed model, using a public dataset of 5000 images for testing and 1500 images for training.

The training dataset was independent from the testing dataset, it was randomly selected from another part of the dataset that was not included in the testing dataset.

The steganalyzer combined two training feature datasets, to deal with the two embedding methods that were used to generate the stego images.

The average of the detection accuracy ranged from 97.50% to 98.73% in the validation test and 97.82% to 98.28% in the field test.

Main Subjects

Information Technology and Computer Science

Topics

• Cryptography

No. of Pages

58

Table of Contents

• APA
• MLA
• AMA

Language

English

Data Type

Arab Theses

Record ID

BIM-762707