Spectral analysis for anomaly detection in dynamic networks with attributes

Detalhes bibliográficos
Autor(a) principal: Silva, Rodrigo Francisquini da [UNIFESP]
Data de Publicação: 2023
Tipo de documento: Tese
Idioma: eng
Título da fonte: Repositório Institucional da UNIFESP
dARK ID: ark:/48912/001300000jnr2
Texto Completo: https://repositorio.unifesp.br/handle/11600/70152
Resumo: Anomaly detection in diverse datasets is a critical area of research with applicability across a wide range of domains, from cybersecurity, such as in intrusion detection in computer networks, to the financial sector, such as in identifying fraudulent activities in credit card transactions. In scenarios where data can be represented as graphs, graph theory offers a set of metrics and methodologies that are particularly effective in capturing the complex relationships and inherent structures within the data in question. However, anomaly detection in graphs presents a series of intricate challenges that have not yet been fully resolved. One such challenge is the dynamic nature of graphs, which evolve over time, rendering static techniques inadequate. Additionally, the presence of heterogeneous attributes on the vertices and edges of the graph increases the complexity of the problem. Traditional methods often fail to adapt to these temporal and spatial changes and frequently lack the interpretability required for real-world applications. To mitigate these challenges, anomaly detection strategies employing clustering techniques have received increasing attention in the literature. These strategies have the advantage of analyzing clusters or groups of vertices, allowing for a more comprehensive and holistic understanding of the underlying graph structure. Such an approach significantly enhances the method's ability to identify not only isolated anomalies but also anomalies that may be indicative of broader structural issues within the graph. The primary objective of this thesis is to investigate and develop unsupervised strategies for anomaly detection in dynamic graphs that also possess heterogeneous attributes. The proposed strategy aims to identify both structural and contextual anomalies. For the detection of structural anomalies, the analysis focuses on the contribution of vertices to the modularity of a specific network partition. On the other hand, contextual anomalies are identified through the application of spectral operators, such as the Fourier Transform. Computational experiments and case studies using real-world datasets corroborate the efficacy of the proposed method. The results demonstrate that the approach outperforms conventional methods found in the literature in terms of both accuracy and interpretability, making it a significant contribution to the field of anomaly detection in graphs.
id UFSP_c6cc9408276ad4fdd1546ccf17168bfc
oai_identifier_str oai:repositorio.unifesp.br/:11600/70152
network_acronym_str UFSP
network_name_str Repositório Institucional da UNIFESP
repository_id_str 3465
spelling Spectral analysis for anomaly detection in dynamic networks with attributesTeoria espectral para detecção de anomalias em redes dinâmicas com atributosAnomaly detectionGraph theorySpectral theoryGraph signal processingAnomaly detection in diverse datasets is a critical area of research with applicability across a wide range of domains, from cybersecurity, such as in intrusion detection in computer networks, to the financial sector, such as in identifying fraudulent activities in credit card transactions. In scenarios where data can be represented as graphs, graph theory offers a set of metrics and methodologies that are particularly effective in capturing the complex relationships and inherent structures within the data in question. However, anomaly detection in graphs presents a series of intricate challenges that have not yet been fully resolved. One such challenge is the dynamic nature of graphs, which evolve over time, rendering static techniques inadequate. Additionally, the presence of heterogeneous attributes on the vertices and edges of the graph increases the complexity of the problem. Traditional methods often fail to adapt to these temporal and spatial changes and frequently lack the interpretability required for real-world applications. To mitigate these challenges, anomaly detection strategies employing clustering techniques have received increasing attention in the literature. These strategies have the advantage of analyzing clusters or groups of vertices, allowing for a more comprehensive and holistic understanding of the underlying graph structure. Such an approach significantly enhances the method's ability to identify not only isolated anomalies but also anomalies that may be indicative of broader structural issues within the graph. The primary objective of this thesis is to investigate and develop unsupervised strategies for anomaly detection in dynamic graphs that also possess heterogeneous attributes. The proposed strategy aims to identify both structural and contextual anomalies. For the detection of structural anomalies, the analysis focuses on the contribution of vertices to the modularity of a specific network partition. On the other hand, contextual anomalies are identified through the application of spectral operators, such as the Fourier Transform. Computational experiments and case studies using real-world datasets corroborate the efficacy of the proposed method. The results demonstrate that the approach outperforms conventional methods found in the literature in terms of both accuracy and interpretability, making it a significant contribution to the field of anomaly detection in graphs.Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP)2017/24185-0Universidade Federal de São PauloNascimento, Mariá Cristina Vasconceloshttp://lattes.cnpq.br/1010810293243435http://lattes.cnpq.br/6809565539156314Silva, Rodrigo Francisquini da [UNIFESP]2024-01-02T17:19:56Z2024-01-02T17:19:56Z2023-11-10info:eu-repo/semantics/doctoralThesisinfo:eu-repo/semantics/publishedVersion136 fapplication/pdfhttps://repositorio.unifesp.br/handle/11600/70152ark:/48912/001300000jnr2enginfo:eu-repo/semantics/openAccessreponame:Repositório Institucional da UNIFESPinstname:Universidade Federal de São Paulo (UNIFESP)instacron:UNIFESP2024-08-13T14:57:47Zoai:repositorio.unifesp.br/:11600/70152Repositório InstitucionalPUBhttp://www.repositorio.unifesp.br/oai/requestbiblioteca.csp@unifesp.bropendoar:34652024-12-11T20:21:25.425596Repositório Institucional da UNIFESP - Universidade Federal de São Paulo (UNIFESP)false
dc.title.none.fl_str_mv Spectral analysis for anomaly detection in dynamic networks with attributes
Teoria espectral para detecção de anomalias em redes dinâmicas com atributos
title Spectral analysis for anomaly detection in dynamic networks with attributes
spellingShingle Spectral analysis for anomaly detection in dynamic networks with attributes
Silva, Rodrigo Francisquini da [UNIFESP]
Anomaly detection
Graph theory
Spectral theory
Graph signal processing
title_short Spectral analysis for anomaly detection in dynamic networks with attributes
title_full Spectral analysis for anomaly detection in dynamic networks with attributes
title_fullStr Spectral analysis for anomaly detection in dynamic networks with attributes
title_full_unstemmed Spectral analysis for anomaly detection in dynamic networks with attributes
title_sort Spectral analysis for anomaly detection in dynamic networks with attributes
author Silva, Rodrigo Francisquini da [UNIFESP]
author_facet Silva, Rodrigo Francisquini da [UNIFESP]
author_role author
dc.contributor.none.fl_str_mv Nascimento, Mariá Cristina Vasconcelos
http://lattes.cnpq.br/1010810293243435
http://lattes.cnpq.br/6809565539156314
dc.contributor.author.fl_str_mv Silva, Rodrigo Francisquini da [UNIFESP]
dc.subject.por.fl_str_mv Anomaly detection
Graph theory
Spectral theory
Graph signal processing
topic Anomaly detection
Graph theory
Spectral theory
Graph signal processing
description Anomaly detection in diverse datasets is a critical area of research with applicability across a wide range of domains, from cybersecurity, such as in intrusion detection in computer networks, to the financial sector, such as in identifying fraudulent activities in credit card transactions. In scenarios where data can be represented as graphs, graph theory offers a set of metrics and methodologies that are particularly effective in capturing the complex relationships and inherent structures within the data in question. However, anomaly detection in graphs presents a series of intricate challenges that have not yet been fully resolved. One such challenge is the dynamic nature of graphs, which evolve over time, rendering static techniques inadequate. Additionally, the presence of heterogeneous attributes on the vertices and edges of the graph increases the complexity of the problem. Traditional methods often fail to adapt to these temporal and spatial changes and frequently lack the interpretability required for real-world applications. To mitigate these challenges, anomaly detection strategies employing clustering techniques have received increasing attention in the literature. These strategies have the advantage of analyzing clusters or groups of vertices, allowing for a more comprehensive and holistic understanding of the underlying graph structure. Such an approach significantly enhances the method's ability to identify not only isolated anomalies but also anomalies that may be indicative of broader structural issues within the graph. The primary objective of this thesis is to investigate and develop unsupervised strategies for anomaly detection in dynamic graphs that also possess heterogeneous attributes. The proposed strategy aims to identify both structural and contextual anomalies. For the detection of structural anomalies, the analysis focuses on the contribution of vertices to the modularity of a specific network partition. On the other hand, contextual anomalies are identified through the application of spectral operators, such as the Fourier Transform. Computational experiments and case studies using real-world datasets corroborate the efficacy of the proposed method. The results demonstrate that the approach outperforms conventional methods found in the literature in terms of both accuracy and interpretability, making it a significant contribution to the field of anomaly detection in graphs.
publishDate 2023
dc.date.none.fl_str_mv 2023-11-10
2024-01-02T17:19:56Z
2024-01-02T17:19:56Z
dc.type.driver.fl_str_mv info:eu-repo/semantics/doctoralThesis
dc.type.status.fl_str_mv info:eu-repo/semantics/publishedVersion
format doctoralThesis
status_str publishedVersion
dc.identifier.uri.fl_str_mv https://repositorio.unifesp.br/handle/11600/70152
dc.identifier.dark.fl_str_mv ark:/48912/001300000jnr2
url https://repositorio.unifesp.br/handle/11600/70152
identifier_str_mv ark:/48912/001300000jnr2
dc.language.iso.fl_str_mv eng
language eng
dc.rights.driver.fl_str_mv info:eu-repo/semantics/openAccess
eu_rights_str_mv openAccess
dc.format.none.fl_str_mv 136 f
application/pdf
dc.publisher.none.fl_str_mv Universidade Federal de São Paulo
publisher.none.fl_str_mv Universidade Federal de São Paulo
dc.source.none.fl_str_mv reponame:Repositório Institucional da UNIFESP
instname:Universidade Federal de São Paulo (UNIFESP)
instacron:UNIFESP
instname_str Universidade Federal de São Paulo (UNIFESP)
instacron_str UNIFESP
institution UNIFESP
reponame_str Repositório Institucional da UNIFESP
collection Repositório Institucional da UNIFESP
repository.name.fl_str_mv Repositório Institucional da UNIFESP - Universidade Federal de São Paulo (UNIFESP)
repository.mail.fl_str_mv biblioteca.csp@unifesp.br
_version_ 1818602473860890624