Privacy Engineering for Governed AI: Protecting High-Dimensional Data in Public and Critical Systems

Abstract

Artificial intelligence depends on large and often high-dimensional datasets, creating privacy risks that are not fully addressed by conventional policy statements. Public agencies and infrastructure operators may process health records, mobility traces, sensor streams, energy-use patterns, images, financial information, and behavioural data. Even when direct identifiers are removed, combinations of attributes can permit re-identification or reveal sensitive facts. This review examines privacy as an engineering, governance, and accountability problem across the AI lifecycle. It is based on the thematic evidence of a systematic synthesis that retained 95 high-quality AI governance studies and frameworks from 2020 to mid-2025. The literature shows that privacy is one of the most frequently recognised governance dimensions, yet implementation quality varies. Many frameworks state requirements for lawful processing and protection but do not define measurable controls for data minimisation, linkage risk, inference attacks, model leakage, access management, and retention. The review compares major privacy controls, including anonymisation, pseudonymisation, differential privacy, secure access, federated approaches, encryption, data lineage, and impact assessment. It also analyses trade-offs between privacy, fairness, transparency, and model utility. A principles-controls-evidence architecture is proposed to make privacy commitments auditable. The article concludes that privacy-preserving AI requires continuous risk assessment, technically appropriate safeguards, and clear institutional responsibility rather than one-time consent or de-identification.