|
|
|
“A Green Learning Approach to Image Forensics: Methodology, Applications, and Performance Evaluation”
by Yao Zhu
August 2023
Fake multimedia has become a central problem in the last few years, especially after the advent of neural networks. Fake multimedia are usually created by whole generation, partial tampering or information hiding. Media forensics, on the contrary, aims to detect the fake contents or discover the hidden information from fake objects. It leverages the fact that manipulation actions leave detectable traces, making fake media objects statistically distinguishable from genuine ones.
In this dissertation, we specifically study two long-standing problems in image forensics: GAN-generated image detection and spatial image steganalysis. The former one aims to detect images that are synthesized by generative models. The latter one focus on distin- guishing stego and cover images in spatial domain, where stego images are generated by various content-adaptive steganography algorithms. The stego signal that are embedded into cover images is so weak that the difference in pixel domain in only +1 or -1.
Existing GAN-generated image detection methods are all based on deep neural networks. However, they often need enormous amount of data or intensive data augmentation to maintain its performance with respect to unseen dataset. This motivates us to find a green (light-weight), robust and high-performance GAN-fake image detector. We propose RGGID, which utilizes the assumption that generative models often fail to synthesize well on high- quality details or complex texture regions. We make decision on blocks from those regions and select discriminant soft scores for image-wise decision fusion. RGGID offers a green solution for GAN-generated image detector since its model size is significantly smaller than that of deep neural networks (DNNs). We apply common manipulations to real/fake source images, including JPEG compression, resizing and Gaussian additive noise, and demonstrate the robustness of RGGID to these manipulations. Furthermore, we prove the generalization ability of RGGID on 11 unseen generative architecture and dataset by training soley on ProGAN and test on other dataset.
Compared to GAN-fake image detection, image steganalysis is a more challenging task. There exist traditional method and deep learning-based method for steganalysis. CNN- based models are proved to have better performance than the three-step traditional machine learning method. However, more secure and complicated steganography schemes force CNN architectures to go deeper and denser, which inevitably results in the insatiable need of mem- ory and computational resources. Motivated by the disadvantages of both methods, we pro- posed a novel learning solution to image steganalysis based on the green learning paradigm, called Green Steganalyzer (GS). GS consists of three modules: 1) pixel-based anomaly pre- diction, 2) embedding location detection, and 3) decision fusion for image-level detection. In the first module, GS decomposes an image into patches, adopts Saab transforms for feature extraction, and conducts self-supervised learning to predict an anomaly score of their center pixel. In the second module, GS analyzes the anomaly scores of a pixel and its neighborhood to find pixels of higher embedding probabilities. In the third module, GS focuses on pixels of higher embedding probabilities and fuses their anomaly scores to make final image-level clas- sification. Compared with state-of-the-art deep-learning models, GS achieves comparable detection performance against S-UNIWARD, WOW and HILL steganography schemes with significantly lower computational complexity and a smaller model size, making it attractive for mobile/edge applications. Furthermore, GS is mathematically transparent because of its modular design supported by logical arguments.