Privacy-Preserving MLOps with Differential Privacy and AI-Guided Tuning

Abstract

When machine learning systems transition from being deployed within research environments to enterprise-scale deployment pipelines, protecting data privacy poses an increasing challenge while the model is being trained and/or used. Privacy-preserving techniques will predominantly rely on some form of static differential privacy (DP) constraint, with the challenge often being to balance privacy requirements with model performance, particularly with dynamic workloads. In this paper, we propose a new Privacy-Preserving MLOps (PP-MLOps) framework that combines AI-aided adaptive tuning of differential privacy in the automated MLOps lifecycle. These proposed agent approaches allow for a flexible way to continuously assess privacy risks, regulatory obligations to privacy and confidentiality, and the value of model utility metrics while adapting DP scale of noise, depth of clipping, and privacy budgets (ε, δ) in real-time to achieve optimal model utility. The continuous optimization of DP in CI/CD pipeline operations is actualized through also using reinforcement-learning based controllers to adjust for a range of privacy and performance tradeoff situations in real-time. Evaluation simulations show a 20% improvement in model accuracy retention in regulation compliant DP tuning and operational measures against traditional fixed DP configurations of varying distributions and operational risks. This study can lay the foundation for fully autonomous risk sensing and regulatory compliant MLOps while translating theoretical claims of privacy, application and assurance in a pragmatic framework machine learning deployment at scale.