Abstract

Given the exponential growth of data in various sectors, the timely identification of anomalies has become crucial in mitigating risks such as financial losses and data breaches. This research focuses on advancing multivariate time series anomaly detection through the innovative application of neuro-evolutionary methods. Utilizing the Evolutionary eXploration of Augmenting Memory Models (EXAMM) algorithm, the study automates the design and optimization of Long Short-Term Memory (LSTM) autoencoder networks, enabling enhanced forecasting capabilities in recurrent neural network architectures. Leveraging these networks along with a One-Class Support Vector Machine (OCSVM) to generate boundary thresholds, this project delivers a robust and scalable procedure for identifying anomalies that outperforms existing proven methods in terms of accuracy and cumulative F1 score, while at the same time requiring orders of magnitude fewer trainable parameters. By comparing the reconstruction error of trained forecasts, the proposed approach facilitates accurate anomaly detection across diverse datasets against unseen anomalous data and informs efficient design decisions for future autoencoder network architectures.

Library of Congress Subject Headings

Anomaly detection (Computer security); Multivariate analysis--Data processing

Publication Date

2025

Document Type

Thesis

Student Type

Graduate

Degree Name

Computer Science (MS)

Department, Program, or Center

Computer Science, Department of

College

Golisano College of Computing and Information Sciences

Advisor

Alexander Ororbia

Advisor/Committee Member

Travis Desell

Advisor/Committee Member

Rui Li

Campus

RIT – Main Campus

Plan Codes

COMPSCI-MS

Share

COinS