Date of Award
Doctor of Philosophy in Electrical Engineering - (Ph.D.)
Electrical and Computer Engineering
Yun Q. Shi
Atam P. Dhawan
Richard A. Haddad
Frank Y. Shih
The security and trustworthiness of digital images have become crucial issues due to the simplicity of malicious processing. Therefore, the research on image steganalysis (determining if a given image has secret information hidden inside) and image forensics (determining the origin and authenticity of a given image and revealing the processing history the image has gone through) has become crucial to the digital society.
In this dissertation, the steganalysis and forensics of digital images are treated as pattern classification problems so as to make advanced machine learning (ML) methods applicable. Three topics are covered: (1) architectural design of convolutional neural networks (CNNs) for steganalysis, (2) statistical feature extraction for camera model classification, and (3) real-world tampering detection and localization.
For covert communications, steganography is used to embed secret messages into images by altering pixel values slightly. Since advanced steganography alters the pixel values in the image regions that are hard to be detected, the traditional ML-based steganalytic methods heavily relied on sophisticated manual feature design have been pushed to the limit. To overcome this difficulty, in-depth studies are conducted and reported in this dissertation so as to move the success achieved by the CNNs in computer vision to steganalysis. The outcomes achieved and reported in this dissertation are: (1) a proposed CNN architecture incorporating the domain knowledge of steganography and steganalysis, and (2) ensemble methods of the CNNs for steganalysis. The proposed CNN is currently one of the best classifiers against steganography.
Camera model classification from images aims at assigning a given image to its source capturing camera model based on the statistics of image pixel values. For this, two types of statistical features are designed to capture the traces left by in-camera image processing algorithms. The first is Markov transition probabilities modeling block-DCT coefficients for JPEG images; the second is based on histograms of local binary patterns obtained in both the spatial and wavelet domains. The designed features serve as the input to train support vector machines, which have the best classification performance at the time the features are proposed.
The last part of this dissertation documents the solutions delivered by the author’s team to The First Image Forensics Challenge organized by the Information Forensics and Security Technical Committee of the IEEE Signal Processing Society. In the competition, all the fake images involved were doctored by popular image-editing software to simulate the real-world scenario of tampering detection (determine if a given image has been tampered or not) and localization (determine which pixels have been tampered). In Phase-1 of the Challenge, advanced steganalysis features were successfully migrated to tampering detection. In Phase-2 of the Challenge, an efficient copy-move detector equipped with PatchMatch as a fast approximate nearest neighbor searching method were developed to identify duplicated regions within images. With these tools, the author’s team won the runner-up prizes in both the two phases of the Challenge.
Xu, Guanshuo, "Machine learning based digital image forensics and steganalysis" (2017). Dissertations. 10.