Author ORCID Identifier

https://orcid.org/0000-0001-6897-0837

Date Available

7-24-2023

Year of Publication

2023

Degree Name

Doctor of Philosophy (PhD)

Document Type

Doctoral Dissertation

College

Engineering

Department/School/Program

Computer Science

First Advisor

Dr. D. Manivannan

Second Advisor

Dr. A. B. Siddique

Abstract

The incidence of computer network intrusions has significantly increased over the last decade, partially attributed to a thriving underground cyber-crime economy and the widespread availability of advanced tools for launching such attacks. To counter these attacks, researchers in both academia and industry have turned to machine learning (ML) techniques to develop Intrusion Detection Systems (IDSes) for computer networks. However, many of the datasets use to train ML classifiers for detecting intrusions are not balanced, with some classes having fewer samples than others. This can result in ML classifiers producing suboptimal results. In this dissertation, we address this issue and present better ML based solutions for intrusion detection. Our contributions in this direction can be summarized as follows:

Balancing Data Using Synthetic Data to detect intrusions in Computer Networks: In the past, researchers addressed the issue of imbalanced data in datasets by using over-sampling and under-sampling techniques. In this study, we go beyond such traditional methods and utilize a synthetic data generation method called Con- ditional Generative Adversarial Network (CTGAN) to balance the datasets and in- vestigate its impact on the performance of widely used ML classifiers. To the best of our knowledge, no one else has used CTGAN to generate synthetic samples for balancing intrusion detection datasets. We use two widely used publicly available datasets and conduct extensive experiments and show that ML classifiers trained on these datasets balanced with synthetic samples generated by CTGAN have higher prediction accuracy and Matthew Correlation Coefficient (MCC) scores than those trained on imbalanced datasets by 8% and 13%, respectively.

Deep Learning approach for intrusion detection using focal loss function: To overcome the data imbalance problem for intrusion detection, we leverage the specialized loss function, called focal loss, that automatically down-weighs easy ex- amples and focuses on the hard negatives by facilitating dynamically scaled-gradient updates for training ML models effectively. We implement our approach using two well-known Deep Learning (DL) neural network architectures. Compared to training DL models using cross-entropy loss function, our approach (training DL models using focal loss function) improved accuracy, precision, F1 score, and MCC score by 24%, 39%, 39%, and 60% respectively.

Efficient Deep Learning approach to detect Intrusions using Few-shot Learning: To address the issue of imbalance the datasets and develop a highly effective IDS, we utilize the concept of few-shot learning. We present a Few-Shot and Self-Supervised learning framework, called FS3, for detecting intrusions in IoT networks. FS3 works in three phases. Our approach involves first pretraining an encoder on a large-scale external dataset in a selfsupervised manner. We then employ few-shot learning (FSL), which seeks to replicate the encoder’s ability to learn new patterns from only a few training examples. During the encoder training us- ing a small number of samples, we train them contrastively, utilizing the triplet loss function. The third phase introduces a novel K-Nearest neighbor algorithm that sub- samples the majority class instances to further reduce imbalance and improve overall performance. Our proposed framework FS3, utilizing only 20% of labeled data, out- performs fully supervised state-of-the-art models by up to 42.39% and 43.95% with respect to the metrics precision and F1 score, respectively.

The rapid evolution of the automotive industry and advancements in wireless com- munication technologies will result in the widespread deployment of Vehicular ad hoc networks (VANETs). However, despite the network’s potential to enable intelligent and autonomous driving, it also introduces various attack vectors that can jeopardize its security. In this dissertation, we present efficient privacy-preserving authenticated message dissemination scheme in VANETs.

Conditional Privacy-preserving Authentication and Message Dissemination Scheme using Timestamp based Pseudonyms: To authenticate a message sent by a vehicle using its pseudonym, a certificate of the pseudonym signed by the central authority is generally utilized. If a vehicle is found to be malicious, certificates associated with all the pseudonyms assigned to it must be revoked. Certificate revocation lists (CRLs) should be shared with all entities that will be corresponding with the vehicle. As each vehicle has a large pool of pseudonyms allocated to it, the CRL can quickly grow in size as the number of revoked vehicles increases. This results in high storage overheads for storing the CRL, and significant authentication overheads as the receivers must check their CRL for each message received to verify its pseudonym. To address this issue, we present a timestamp-based pseudonym allocation scheme that reduces the storage overhead and authentication overhead by streamlining the CRL management process.

Digital Object Identifier (DOI)

https://doi.org/10.13023/etd.2023.292

Share

COinS