Research on AI Security Strategies and Practical Approaches for Risk Management

Abstract

The rapid integration of Artificial Intelligence (AI) into critical infrastructures has exposed significant security vulnerabilities that traditional cybersecurity paradigms fail to address. Unlike deterministic IT systems, AI models are stochastic and data-dependent, making them susceptible to unique threats such as data poisoning, adversarial evasion, and model inversion. This dissertation investigates the comprehensive landscape of AI security, aiming to bridge the gap between isolated technical defenses and holistic organizational risk management. Through a systematic review and taxonomic analysis, this research categorizes AI risks across three distinct layers: data integrity, model robustness, and system deployment. The study evaluates current defense strategies, demonstrating that technical measures like adversarial training and differential privacy are necessary but insufficient when applied in isolation. Consequently, a practical AI Risk Management Framework is proposed, structured around a continuous lifecycle of mapping, measuring, managing, and monitoring risks. The findings suggest that effective AI security requires a "Defense-in-Depth" strategy that integrates robust MLOps infrastructure with rigorous governance policies. The dissertation concludes that shifting from static security controls to dynamic, lifecycle-based assurance is essential for deploying trustworthy AI systems in high-stakes environments.