Performance evaluation of serverless cloud-native API deployment: a case study on a mobile health application

Abstract

The increasing integration of artificial intelligence (AI) in cybersecurity has enhanced the ability to detect and mitigate cyber threats in real-time. However, adversarial machine learning (AML) has emerged as a significant challenge, enabling attackers to manipulate AI models and bypass security measures. This study explores the evolving landscape of AML threats and the vulnerabilities they introduce to AI-powered defense systems. The research identifies key adversarial attack techniques, including evasion, poisoning, model inversion, and model extraction, which threaten the integrity and effectiveness of AI-driven cybersecurity mechanisms. This study evaluates various mitigation strategies to address these threats, such as adversarial Training, model hardening, defensive Distillation, and hybrid AI approaches. Through experimental analysis, we assess the robustness of AI defense systems under adversarial attack and measure their effectiveness using key performance metrics, including model accuracy, false positive rates, and computational efficiency. The findings indicate that while adversarial Training improves model resilience, adaptive attack techniques continue to challenge existing defenses, necessitating continuous advancements in cybersecurity frameworks. This research highlights the need for a multi-layered security approach that integrates AI-based anomaly detection, human-AI hybrid security models, and adaptive learning techniques to counter adversarial threats effectively. Additionally, it discusses the broader implications of AML in cybersecurity, including policy considerations, ethical concerns, and future research directions. The study recommends strategies for enhancing AI-powered cyber defense systems to maintain security, reliability, and resilience against evolving adversarial threats.