Study on Risk Assessment Methods and Multi-Dimensional Control Mechanisms in AI Systems

As Artificial Intelligence (AI) rapidly transitions from experimental prototypes to critical infrastructure, the historical "Performance-First" paradigm has left systems inherently vulnerable to adversarial attacks and data manipulation. This dissertation addresses the critical lack of standardized, quantitative methods for managing these risks by introducing the Risk Assessment Model for AI (RAM-AI). Utilizing a dual-domain simulation approach across Computer Vision and Financial datasets, the study empirically quantifies the "robustness boundary" of deep learning models. The findings reveal that single-layer defenses are inadequate; specifically, models exhibit "Data Hypersensitivity," suffering non-linear performance collapse under data poisoning rates as low as 3%. Furthermore, standard accuracy metrics fail to detect high-confidence evasion attacks. To mitigate these vulnerabilities, the research validates a Multi-dimensional Control Framework that integrates technical safeguards-such as adversarial training and input sanitization-with procedural governance, including Human-in-the-Loop (HITL) protocols. The results demonstrate that this Defense-in-Depth architecture significantly recovers system integrity, reducing critical error rates by 88% in high-stakes scenarios, and offers a strategic playbook for Enterprise Risk Management in the era of emerging AI regulations.