arXiv: Lifecycle-Aware Dynamic Analysis for Secure ML Model Execution

This publication introduces a technical framework for lifecycle-aware dynamic analysis of machine learning models, aimed at detecting and mitigating security vulnerabilities during execution. The authors propose a method that monitors model behavior across its entire lifecycle, from training to deployment, to identify threats such as adversarial attacks, data poisoning, and model inversion. While not a regulatory mandate, this research signals an emerging best practice for secure ML model execution under the AI_SAFETY framework, which is increasingly referenced in EU AI Act compliance discussions.

The primary affected organizations are those developing or deploying high-risk AI systems, particularly in finance, healthcare, critical infrastructure, and cloud-based ML platforms. Compliance teams in these sectors should note that dynamic, lifecycle-aware security analysis may soon become an expected component of risk management and conformity assessments. Regulators are likely to view such proactive monitoring as evidence of robust technical safeguards.

Compliance teams should immediately review their current ML model validation and monitoring procedures against this lifecycle-aware approach. Begin by mapping your models’ security controls across training, validation, and deployment phases. Engage with technical teams to assess feasibility of integrating dynamic analysis tools, and document any gaps for future audit readiness. This publication does not change current legal obligations, but it provides a clear technical benchmark for demonstrating compliance with AI safety and security requirements.