Real-Time Machine Learning for Quickest Detection

<p>Safety-critical Cyber-Physical Systems (CPS) require real-time machine learning for control and decision making. One promising solution is to use deep learning to discover useful patterns for event detection from heterogeneous data. However, deep learning algorithms encounter challenges in CPS with assurability requirements: 1) Decision explainability, 2) Real-time and quickest event detection, and 3) Time-eficient incremental learning.</p> <p>To address these obstacles, I developed a real-time Machine Learning Framework for Quickest Detection (MLQD). To be specific, I first propose the zero-bias neural network, which removes decision bias and preferabilities from regular neural networks and provides an interpretable decision process. Second, I discover the latent space characteristic of the zero-bias neural network and the method to mathematically convert a Deep Neural Network (DNN) classifier into a performance-assured binary abnormality detector. In this way, I can seamlessly integrate the deep neural networks' data processing capability with Quickest Detection (QD) and provide real-time sequential event detection paradigm. Thirdly, after discovering that a critical factor that impedes the incremental learning of neural networks is the concept interference (confusion) in latent space, and I prove that to minimize interference, the concept representation vectors (class fingerprints) within the latent space need to be organized orthogonally and I invent a new incremental learning strategy using the findings, I facilitate deep neural networks in the CPS to evolve efifciently without retraining. All my algorithms are evaluated on real-world applications, ADS-B (Automatic Dependent Surveillance Broadcasting) signal identification, and spoofing detection in the aviation communication system. Finally,</p> <p>I discuss the current trends in MLQD and conclude this dissertation by presenting the future research directions and applications.</p> <p>As a summary, the innovations of this dissertation are as follows: i) I propose the zerobias neural network, which provides transparent latent space characteristics, I apply it to solve the wireless device identification problem. ii) I discover and prove the orthogonal memory organization mechanism in artificial neural networks and apply this mechanism in time-efficient incremental learning. iii) I discover and mathematically prove the converging point theorem, with which we can predict the latent space topological characteristics and estimate the topological maturity of neural networks. iv) I bridge the gap between machine learning and quickest detection with assurable performance.</p>

Name thesis:degree_name

Doctor of Philosophy in Electrical Engineering & Computer Science

Level thesis:degree_level

Dissertation - Open Access

Discipline thesis:degree_discipline

Electrical Engineering and Computer Science

Year

2021

Author dc:creator

• Liu, Yongxin

• machine learning
• detection
• Cyber-Physical Systems
• cybersecurity
• Computer Engineering
• Electrical and Computer Engineering

Repository record dc:identifier

https://commons.erau.edu/edt/614

OAI identifier oai:identifier

oai:commons.erau.edu:edt-1625