If you have any problems related to the accessibility of any content (or if you want to request that a specific publication be accessible), please contact us at scholarworks@unr.edu.
Robust and cross-domain anomaly detection and mitigation
Date
2024Type
DissertationDepartment
Computer Science
Degree Level
Doctorate Degree
Abstract
Anomaly detection aims to identify unusual patterns in data that significantly diverge from normal instances. When an anomaly is detected, specific steps are undertaken to address and resolve the issue. Despite significant advancements inanomaly detection techniques in recent years, challenges such as low recall rates, extreme class imbalance, and high noise levels persist. The success of mitigation closely depends on the detection phase. As a strategy to enhance anomaly detection, developing a system that creates an ideal environment for this purpose is proposed. In the realm of supply chain security, to counteract anomaly attacks and enhance mitigation, we suggest the ’3D Unclonable Optical Identity’ as a solution for product verification. This tag, designed with a distinctive 3D structure, is exceedingly difficult to replicate, even with advanced fabrication methods. The security of this ID rests on the difficulty of duplication, rather than on secrecy. To address issues such as class imbalance and the impact of noise in practical applications, we utilized UE4 to generate thousands of simulated images from different angles and lighting conditions. This method assists in the development of an anomaly detection system capable of identifying counterfeit tags.Another challenge in anomaly detection is identifying anomalies across various data types, necessitating a versatile, data-agnostic method for representing typical samples. Because of these challenges, specific models are frequently developed fordifferent anomaly detection applications. A possible solution is to employ a single model to detect diverse types of anomalies. The generative model, especially the diffusion model, has attracted attention due to its ability to create high-quality images
and potential in enhancing anomaly detection. We propose a latent diffusion-based multi-class anomaly detection model. This model learns latent representations of non-anomalous samples and is capable of detecting anomalies in multiple classes.
Our extensive evaluations on benchmark datasets such as MNIST and CIFAR-10 have shown that our approach outperforms current state-of-the-art methods in latent diffusion-based anomaly detection.Anomaly detection in the biomedical imaging field presents unique challenges, chiefly in accurately segmenting anomaly areas and quantifying anomaly behaviors. Gould Syndrome, a rare genetic multi-system disorder, is one such case. We havedeveloped a Gould Syndrome Detection pipeline to detect gene changes based on vascular SMC phenotype. Additionally, calcium imaging, a crucial regulatory mechanism for cerebral blood flow, is addressed in our work. We have created SEANVC
(Simple Semi-automated Analytical Tool for Astrocyte Ca2+ Signals and Vascular Responses in Neurovascular Coupling) to assist researchers in identifying anomalous Ca2+ signals and their corresponding vascular responses.
Permanent link
http://hdl.handle.net/11714/12709Additional Information
Committee Member | Bebis, George; Nicolescu, Mircea; Sengupta, Shamik; Murray, Nicholas |
---|