Privacy-Preserving Learning Techniques for Responsible AI

Abstract

Privacy-preserving machine learning (PPML) -- encompassing techniques that enable AI model training and inference while providing formal guarantees against the disclosure of individual training data -- has emerged as a critical enabler of responsible AI deployment in domains where training data contains sensitive personal information. The EU AI Act (2024) and GDPR (2016) jointly create a regulatory environment in which privacy is not merely a compliance obligation but an architectural design requirement for high-risk AI systems processing personal data. Despite substantial technical advances in differential privacy (DP), federated learning (FL), secure multi-party computation (SMPC), and homomorphic encryption (HE), the practical deployment of PPML in responsible AI systems remains constrained by poorly understood trade-offs among privacy strength, model accuracy, computational overhead, and fairness. This paper presents a systematic comparative evaluation of six PPML techniques across four responsible AI deployment dimensions -- privacy guarantee strength, accuracy-privacy trade-off, computational feasibility, and fairness impact -- applied to three benchmark domains: clinical tabular data, financial time-series data, and medical image classification. Results demonstrate that no single PPML technique dominates across all four dimensions: DP-SGD achieves the strongest privacy guarantees (epsilon = 1.0) but the highest accuracy cost (mean AUC reduction = 5.8%) and the most adverse fairness impact (mean DPD increase = 0.042). Federated learning with secure aggregation achieves strong privacy with moderate accuracy cost (mean AUC reduction = 2.4%) and minimal fairness impact. The paper contributes a domain-stratified PPML evaluation benchmark, a Privacy-Accuracy-Fairness (PAF) trade-off framework for technique selection, and practical guidance for GDPR-compliant and EU AI Act-aligned PPML deployment.