Dynamic anomaly detection in sensor networks

Rovatsos, Georgios

Permalink

https://hdl.handle.net/2142/109408 Copy

Description

Title

Dynamic anomaly detection in sensor networks

Author(s)

Rovatsos, Georgios

Issue Date

2020-12-01

Director of Research (if dissertation) or Advisor (if thesis)

Veeravalli, Venugopal V

Doctoral Committee Chair(s)

Veeravalli, Venugopal V

Committee Member(s)

• Do, Minh N
• Dominguez-Garcia, Alejandro D
• Fellouris, Georgios

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Date of Ingest

2021-03-05T21:38:14Z

Keyword(s)

• Sensor networks
• quickest change detection
• dynamic anomalies
• optimal tests
• stopping times.

Abstract

In the problem of quickest change detection, a sequence of random variables is observed sequentially by a decision maker. At some unknown time instant, the emergence of an anomaly leads to a change in the distribution of the observations. The goal in quickest change detection is to detect this change as quickly as possible, subject to constraints on the frequency of false alarm events. One important application of the theory of quickest change detection is in the context of anomaly detection in sensor networks used to monitor engineering systems. Sensor network related detection problems can vary significantly depending on the spatial evolution of the anomaly in the network as time progresses. Settings involving static anomalies, i.e., anomalies that are perceived by all sensors concurrently and that affect sensors persistently, have been extensively studied in the quickest change detection literature. In addition, semi-dynamic quickest change detection settings that involve anomalies that affect sensors at different time instants, albeit in a persistent manner, have recently received more attention. In this dissertation, our goal is to study the problem of dynamic anomaly detection in sensor networks, i.e., the case where anomalies may not affect sensors persistently, but may move around the network affecting different sets of sensors with time. The objective is to design anomaly detection procedures that are provably optimal with respect to delay-false alarm trade-off formulations. We study the quickest dynamic anomaly detection problem under multiple settings by imposing different assumptions on the spatial evolution of the anomaly. In particular, we consider the case where anomalies evolve according to a discrete-time Markov chain model, for which we develop asymptotically optimal procedures which we compare with more computationally feasible heuristic detection algorithms that require less model knowledge. The Markov model definition incorporates anomalies the size of which may be constant or vary with time. In addition, we study the worst-path dynamic anomaly detection setting, where we assume that the trajectory of the anomaly is unknown and deterministic, and that candidate detection procedures are evaluated according to the anomaly path that maximizes their detection delay. We consider the worst-path setting under the assumption that the anomaly affects a fixed size of sensors, as well as study the problem of worst-path anomaly detection when the size of the anomaly changes with time. For the two worst-path settings we establish that algorithms from quickest change detection literature can be modified to result in provably asymptotically optimal, and in some cases, exactly optimal procedures. A detailed performance analysis of the proposed algorithms is conducted, and concise guidelines regarding the design of proposed tests are provided. Numerical studies of the proposed detection schemes are presented for all studied settings and for a variety of test cases, such as different network sizes, probability distributions, and degrees of model knowledge. Finally, we outline problems of interest for future work, such as the extension of proposed algorithms and techniques in settings where model knowledge is limited.

Graduation Semester

2020-12

Type of Resource

Thesis

Permalink

http://hdl.handle.net/2142/109408

Copyright and License Information

Copyright 2020 Georgios Rovatsos

Owning Collections