Integration of security testing into QA pipelines using adversarial ML

The study aimed to theoretically systematise approaches to improving the resilience of machine learning systems in cyber defence by integrating resilience testing into the security process. The methodology covered the systematisation of machine learning areas in cyber defence, analysis of strategies to counter adversarial attacks, and a case study of integration into quality assurance and machine learning operations. The study found that the use of machine learning technologies in cyber defence enables the automation of threat detection and response (network anomalies, behavioural analysis, anti-phishing, anti-fraud, malware classification). The main advantages are scalability, response speed, predictability, and effectiveness in complex environments, while the key risks include dependence on data quality, false positives, vulnerability to adversarial and poisoning attacks, as well as privacy and explainability issues. The study determined that adversarial machine learning distinguishes between three attack scenarios (white-box, black-box, grey-box) and their classes (evasion, data poisoning, privacy/inference, model extraction, generative artificial intelligence. The study emphasised that adversarial machine learning encompasses not only technical but also regulatory and ethical dimensions related to the principles of privacy, fairness, and transparency in the use of artificial intelligence. Multi-level protection strategies were presented, integrated into the machine learning model lifecycle at the data level, during training, after training, at the deployment and inference stages. Practical cases demonstrated the feasibility of applying machine learning and anti-money laundering in various domains, from network security and security operations centres to development and operations/continuous integration/continuous delivery, the financial sector, stress testing machine learning pipelines, as well as quality assurance and machine learning operations. The practical significance lies in the ability of cybersecurity specialists, financial analysts, and machine learning operations engineers to use the results to improve the efficiency of security operations centres, integrate adversarial testing, and ensure the stability of machine learning models in production environments