Differential privacy limits the exposure of individual data points when an AI system is trained, queried, or compromised. It is one safeguard within a broader privacy-by-design approach, not a substitute for encryption, isolation, or access controls. section 11.2.3 section 11.1.3
The practical goal is to reduce the damage caused by a privacy failure while preserving useful system behavior. This matters because exposed chat histories, billing details, or training data can erode customer trust even when the original failure is technical. section 11.2.3
Use differential privacy as one part of a layered governance decision:
- Apply it when protecting individual data remains important even if a model is stolen or compromised. section 11.1.3
- Pair it with end-to-end encryption, data isolation, and stricter access controls rather than treating it as complete privacy protection. section 11.2.3
- Explain its role alongside operational safeguards such as API-rate limits, watermarking, licensing, and secure data handling. section 11.1.3
Model compromise
A recommendation API may be queried repeatedly to clone its model. Differential privacy helps protect individual training data if the model is compromised, while rate limits and other controls address extraction and unauthorized use. section 11.1.3
