Privacy Auditing

Privacy auditing studies how to measure what an adversary can learn about people from training data, released models, synthetic data, and AI system interactions. Common approaches include membership-inference and reconstruction attacks, optimized canaries, re-identification tests, and statistical tests that estimate privacy loss or certify risk; differential privacy provides a formal reference point. Audits increasingly examine realistic threat models, vulnerable subgroups, and information flows across tools and system components, while balancing privacy findings against utility. A central challenge is that attack results depend on assumptions, data, and evaluation choices, so a failed attack rarely proves that information cannot leak.