Date of Graduation

1-2021

Document Type

Thesis

Degree Name

Master of Science in Computer Engineering (MSCmpE)

Degree Level

Graduate

Department

Computer Science & Computer Engineering

Advisor

H. Alan Mantooth

Committee Member

Qinghua Li

Second Committee Member

Chris Farnell

Third Committee Member

Jia Di

Keywords

grid-connected devices, cyber attacks, recurrent neural network, LSTM autoencoder, intrusion detection models

Abstract

The purpose of this thesis is to develop a reference design for a base level implementation of an intrusion detection module using artificial neural networks that is deployed onto an inverter and runs on live data for cybersecurity purposes, leveraging the latest deep learning algorithms and tools. Cybersecurity in the smart grid industry focuses on maintaining optimal standards of security in the system and a key component of this is being able to detect cyberattacks. Although researchers and engineers aim to design such devices with embedded security, attacks can and do still occur. The foundation for eventually mitigating these attacks and achieving more robust security is to identify them reliably. Thus, a high-fidelity intrusion detection system (IDS) capable of identifying a variety of attacks must be implemented. This thesis provides an implementation of a behavior-based intrusion detection system that uses a recurrent artificial neural network deployed on hardware to detect cyberattacks in real time. Leveraging the growing power of artificial intelligence, the strength of this approach is that given enough data, it is capable of learning to identify highly complex patterns in the data that may even go undetected by humans. By intelligently identifying malicious activity at the fundamental behavior level, the IDS remains robust against new methods of attack. This work details the process of collecting and simulating data, selecting the particular algorithm, training the neural network, deploying the neural network onto hardware, and then being able to easily update the deployed model with a newly trained one. The full system is designed with a focus on modularity, such that it can be easily adapted to perform well on different use cases, different hardware, and fulfill changing requirements. The neural network behavior-based IDS is found to be a very powerful method capable of learning highly complex patterns and identifying intrusion from different types of attacks using a single unified algorithm, achieving up to 98% detection accuracy in distinguishing between normal and anomalous behavior. Due to the ubiquitous nature of this approach, the pipeline developed here can be applied in the future to build in more and more sophisticated detection abilities depending on the desired use case.

The intrusion detection module is implemented in an ARM processor that exists at the communication layer of the inverter. There are four main components described in this thesis that explain the process of deploying an artificial neural network intrusion detection algorithm onto the inverter: 1) monitoring and collecting data through a front-end web based graphical user interface that interacts with a Digital Signal Processor that is connected to power-electronics, 2) simulating various malicious datasets based on attack vectors that violate the Confidentiality-Integrity-Availability security model, 3) training and testing the neural network to ensure that it successfully identifies normal behavior and malicious behavior with a high degree of accuracy, and lastly 4) deploying the machine learning algorithm onto the hardware and having it successfully classify the behavior as normal or malicious with the data feeding into the model running in real time. The results from the experimental setup will be analyzed, a conclusion will be made based upon the work, and lastly discussions of future work and optimizations will be discussed.

Share

COinS