PRIV-ML: Analyzing Privacy Loss in Iterative Machine Learning with Differential Privacy

Differential privacy offers rigorous protections for emerging paradigms like federated machine learning, decentralized analytics, and web3 applications. The parameters E (epsilon) and <5 (delta) are crucial in balancing privacy and utility by bounding the maximum divergence between outputs on neighboring datasets. However, quantifying cumulative privacy loss over long-running processes involving iterative model queries, validations, tuning, and multi-party computations remains an open challenge restricting adoption. This paper proposes a comprehensive methodology to evaluate end-to-end differential privacy guarantees across complex Machine Learning (ML) workflows. We develop a heuristic algorithm that maintains a privacy budget depleted per operation based on computed data sensitivity and noise calibration. By tracking tight stochastic bounds on the cumulative privacy loss random variable using advanced composition theorems, our approach can formally verify guarantees over iterative workflows. Simulations demonstrate the technique quantifying privacy loss across 950 successive histogram queries under (E = 1, <5 = 10-
⁵
)-differential privacy while sustaining utility with an average error of only 4.5% compared to non-private histograms, underscoring the importance of formally tracking cumulative privacy loss. Our framework provides a practical solution for measurable privacy-preserving machine learning pipelines without degrading accuracy or utility. By interfacing with diverse mechanisms and adapting noise to empirical sensitivities, we facilitate precise reasoning of privacy risks throughout model life cycles. We also analyze privacy parameter implications across application domains. This paper lays a rigorous foundation for developing trustworthy AI systems that protect sensitive data.