Date of Award

Fall 12-2025

Document Type

Doctoral Project

Degree Name

Doctor of Engineering (D Eng)

Department

Electrical & Computer Engineering

Program/Concentration

Electrical and Computer Engineering - Cybersecurity

Committee Director

Chen Chung

Committee Director

C. Ariel Pinto

Committee Member

Oscar Gonzales

Abstract

This doctoral project aims to bridge the gap between graph theory and network science to identify and mitigate cyber risk, represented as a CY-Triangular Network that connects different networks. The CY-Triangular Framework is a cybersecurity system that integrates graph theory and network science through an interoperable learning approach. The objective of this project is to bridge the gap between two domains: network science and network systems. Accordingly, it examines one representative network from each field, focuses on a complex system network, and explores Graph Neural Networks (GNNs). The connection between these domains lies in graph theory. This research demonstrates that both network types employ graph theory and are susceptible to cyberattacks, with particular emphasis on backdoor attacks. To support this, we conduct studies on GNNs and complex systems by implementing a backdoor attack and mitigating its effects using both a newly developed ECSMT framework for GNNs and an optimization-based cybersecurity mitigation model, OBCSMT, for complex systems. Results show improved model performance and reduce associated costs.

In this paper, we focus more on the limitations of defense mechanisms for backdoor attacks in GNNs. This can be explained by other challenges, including their failure to properly handle various attack types, their inability to easily identify poisoned nodes, and challenges in generalizing across diverse graph structures. The computational overhead and chances of false positives and negatives also make it harder for them to be effective. To illustrate the limitations of defense mechanisms against backdoor attacks in GNNs. We develop an Explainable CS-Mitigation Triangular (ECSMT) Framework that is composed of three mitigation techniques (Robust Training, Graph Regularization, and Data Sanitization). We implement one or more of the framework’s mitigation techniques in the GNN model to improve performance. We implemented an OBCSMT model across the components of a complex system, reducing total costs by 36% on average, and applied it to specific nodes of the attack graphs, resulting in a 61% reduction in total costs. We applied our novel ECSMT framework to enhance accuracy and reduce the attack success rate of the first-invented backdoor mitigation method, GCleaner, on GNNs. Our research aims to extend the GCleaner framework by integrating it into the ECSMT framework. This study demonstrates that ECSMT techniques can reduce the ASR while maintaining or slightly improving accuracy. Unlike GCleaner, which focuses on trigger recovery and unlearning, our approach includes three additional mitigation strategies from the ECSMT taxonomy: Robust Training, Graph Regularization, and Data Sanitization. These strategies are applied during the training phase, followed by the GCleaner unlearning stage. Quantitative results show that this extended framework achieves low ASR and more stable clean accuracy than unlearning alone. Therefore, our contribution is to extend the GCleaner framework to significantly strengthen the model's defense capabilities. The code uploaded to the GitHub Repository ( https://github.com/Sabouha-17/ECSMT-Framework.git) illustrates how to quantify the effectiveness of our framework against backdoor attacks on GNNs.

Rights

In Copyright. URI: http://rightsstatements.org/vocab/InC/1.0/ This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s).

DOI

10.25777/fehx-gv68

Download

Share

COinS