Privacy Attacks—Extracting Secrets from AI Models

This chapter explores the rapidly evolving threat of privacy attacks on machine learning systems, focusing on how adversaries exploit model behavior to extract sensitive information. It presents a detailed examination of techniques such as membership inference, model inversion, attribute and property inference, and model extraction, revealing how seemingly harmless queries can expose training data and private attributes. Special attention is given to the vulnerabilities of generative models, where prompt-based attacks can coax out memorized sequences and confidential completions. Readers will gain a practical understanding of how attackers use black-box and gray-box access to reconstruct data, replicate model functionality, and infer hidden patterns without breaching traditional system defenses. The chapter explains how output confidence, embedding similarities, and response variability are leveraged as leakage channels, and why high-capacity models trained on unfiltered data are particularly prone to unintentional disclosure. Realistic adversarial workflows are explored, showing how iterative prompting and surrogate modeling fuel increasingly precise extractions. The chapter closes with actionable defensive strategies for artificial intelligence (AI) practitioners and security professionals, emphasizing architectural awareness, inference monitoring, prompt auditing, and posttraining evaluation. Readers will come away equipped to rethink how they assess, deploy, and govern machine learning models from a privacy-resilient standpoint. By reframing model inference as an active risk surface, the chapter positions privacy attacks as a frontline concern in AI system security and compliance.