Realistic adversarial machine learning to improve network intrusion detection

• Autor(a) principal: Vitorino, João Pedro Machado
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.22/23426
• topic: Realistic adversarial examples; Adversarial robustness; Tabular data; Machine learning; Cybersecurity; Exemplos adversos realistas; Robustez adversa; Dados tabulares; Aprendizagem automática; Cibersegurança

Modern organizations can significantly benefit from the use of Artificial Intelligence (AI), and more specifically Machine Learning (ML), to tackle the growing number and increasing sophistication of cyber-attacks targeting their business processes. However, there are several technological and ethical challenges that undermine the trustworthiness of AI. One of the main challenges is the lack of robustness, which is an essential property to ensure that ML is used in a secure way. Improving robustness is no easy task because ML is inherently susceptible to adversarial examples: data samples with subtle perturbations that cause unexpected behaviors in ML models. ML engineers and security practitioners still lack the knowledge and tools to prevent such disruptions, so adversarial examples pose a major threat to ML and to the intelligent Network Intrusion Detection (NID) systems that rely on it. This thesis presents a methodology for a trustworthy adversarial robustness analysis of multiple ML models, and an intelligent method for the generation of realistic adversarial examples in complex tabular data domains like the NID domain: Adaptative Perturbation Pattern Method (A2PM). It is demonstrated that a successful adversarial attack is not guaranteed to be a successful cyber-attack, and that adversarial data perturbations can only be realistic if they are simultaneously valid and coherent, complying with the domain constraints of a real communication network and the class-specific constraints of a certain cyber-attack class. A2PM can be used for adversarial attacks, to iteratively cause misclassifications, and adversarial training, to perform data augmentation with slightly perturbed data samples. Two case studies were conducted to evaluate its suitability for the NID domain. The first verified that the generated perturbations preserved both validity and coherence in Enterprise and Internet-of Things (IoT) network scenarios, achieving realism. The second verified that adversarial training with simple perturbations enables the models to retain a good generalization to regular IoT network traffic flows, in addition to being more robust to adversarial examples. The key takeaway of this thesis is: ML models can be incredibly valuable to improve a cybersecurity system, but their own vulnerabilities must not be disregarded. It is essential to continue the research efforts to improve the security and trustworthiness of ML and of the intelligent systems that rely on it.