News

Congratulations to Yao Zhu for passing her defense on June 12, 2023. Yao’s thesis is titled “A Green Learning Approach to Image Forensics: Methodology, Applications, and Performance Evaluation.” Her Dissertation Committee includes Jay Kuo (Chair), Antonio Ortega, and Jernej Barbic (Outside Member). Yao received several questions and suggestions from the Committee members. Yao answered the questions professionally.

Congratulations to Yao for this milestone moment in life. MCL News team invited Yao for a short talk on her thesis and PhD experience, and here is the summary. We thank Yao for her kind sharing, and wish her all the best in the next journey.

“Fake images have become a central problem in the last few years, especially after the advent of neural networks. Fake images are usually created by whole generation, partial tampering or information hiding. Image 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 images statistically distinguishable from genuine ones.

I specifically talked about 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 distinguishing 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. The solutions that we propose to these two problems are both ‘green’ solutions, which have significantly small model sizes and computational cost. In the meantime, our methods are mathematically transparent due to the modularized design. Green [...]

Camouflage is an attempt to mask the object into a background image and to match its background. The term “camouflage” originates in the ancient practices of the animal kingdom, where animals would alter their body patterns, textures, and colors to blend in with their environment to evade predators. In military contexts, camouflage is a technique used to conceal soldiers or equipment within the background texture, making it difficult for the enemy to detect them. On the other hand, camouflaged object detection is a method used to uncover the enemy who has used camouflage to hide within the image texture.

Our research objectives are to understand the problem with more insights and develop a pipeline for Camouflaged Object Detection based on the Green Learning framework. The first main technical accomplishments are multiscale color and texture decomposition; since, in this task, the texture is much more important than the color, we decouple the information of textures and colors. The second main technical accomplishment is clustering the images dataset; we cluster similar photos with the matching color histogram together and try to learn the information. It reduces the difficulty of the classifier to separate all kinds of images, and we can focus more on images with similar colors and appearance. We cluster the dataset using HSV color space, quantize the color space into 52 bins, and use K-means clustering to derive the clusters. As shown in Figure, images with similar backgrounds are grouped together. Our experiment shows that training in each subcluster could further reduce training and validation loss.

— Max Chen

Martin Ohm (May 6, 1792 in Erlangen – April 1, 1872 in Berlin) was a German mathematician and a younger brother of physicist Georg Ohm. He earned his doctorate in 1811 at Friedrich-Alexander-University, Erlangen-Nuremberg where his advisor was Karl Christian von Langsdorf. In 1817, he was appointed professor of mathematics and physics in the gymnasium at Thorn. In 1821 he moved to Berlin, and in 1839 became a full professor in the University of Berlin. He delivered courses of lectures at the academy of architecture from 1824 to 1831, and at the schools of artillery and engineering from 1833 to 1852; and he also taught in the military school from 1826 to 1849. Ohm was the first to fully develop the theory of the exponential ab when both a and b are complex numbers in 1823. The 1835 second edition of Ohm’s textbook, Die reine Elementar Mathematik was the first time that Euclid’s ‘extreme and mean ratio’ was given the name of the “golden section” (goldener Schnitt). Ohm’s notable students included Rudolf Lipschitz.

— Xiou Ge

Source: https://en.wikipedia.org/wiki/Martin_Ohm

Let us hear what he has to say about his defense and an abstract of his thesis.

Knowledge graphs (KGs) find rich applications in knowledge management and discovery, recommendation systems, fraud detection, chatbots, etc. KGs are directed relational graphs. They are formed by a collection of triples in the form of (head, relation, tail), where heads and tails are called entities and represented by nodes while relations are links in KGs. Since KGs are often incomplete, one critical task in KG is to predict missing information based on available knowledge. Knowledge graph embedding (KGE) is one of the most popular approaches for link prediction and entity typing due to its effectiveness, efficiency, and explainability.

In this thesis, we propose a KGE inspired by compounding three-dimensional geometric transformation named CompoundE3D. Specifically, we apply an adapted beam search algorithm to discover the best performing variants and experiment with different ensemble strategies to achieve state-of-the-art results on four benchmarking datasets. We also analyze the properties of different operators under the CompoundE3D framework and empirically verify their advantages.

I’d like to thank Prof. Kuo for his advising throughout the last 4 years. The working style in MCL has been very effective. Without the weekly reports and meetings, it would probably take me more time to grow. I really respect Prof. Kuo’s research vision and the research direction we embark on will be appreciated in the long term. I’m also very fortunate to be with a group of very nice fellow students in the lab. Everyone is so supportive and willing to help each other. MCL is a perfect fit for students like me who may not be the most talented but are determined to work hard and do good research.

— Xiou [...]

Steganography for spatial images is a technique that involves hiding secret information within digital images in a way that is imperceptible to the human eye. It considers the characteristics and features of the cover media in a local region, and embed the secret information in a manner that is visually or statistically inconspicuous.

As the other side of the coin, steganalysis, is the way of detecting the embedded image (often we call it stego image). There has been numerous works in detecting the hidden information. Traditionally, people use heuristic features and ensemble of machine learning models for detection. It soon becomes feeble in detecting content adaptive steganographic images. After the emerging of neural networks, researchers start to use deep models to detect the weak stego signals by extracting features and doing classification in a whole. Different from the traditional steganalysis and deep learning based steganalysis methods, we propose a novel steganalysis scheme, which is a green steganalysis method.

We first scatter the whole image into small blocks, and then perform anomaly detection on block levels. This step will give us an indication of the likelihood such that this block is embedded or not. Next, we train an embed location detector, to help us locate the blocks that are more discriminant than others. Finally, blocks that are selected from previous step will be fused together and make image-level decision by ensemble classifier. Our architecture is completely explainable, and computationally efficient.

— by Yao Zhu