Differential privacy introduces carefully calibrated noise during training or query responses so that it becomes mathematically difficult to infer whether any specific individual's record is included in the training set. For healthcare data---highly sensitive and subject to strict privacy laws---this technique directly supports privacy-by-design, reducing the risk that model outputs leak membership information or reconstruct personal records.

Option A uses real patient records directly and does not, by itself, mitigate inference risk. Option B (data augmentation) may expand the dataset but does not guarantee resistance to membership inference attacks. Option D (transfer learning using public data) can help, but if any private data is used in fine-tuning, privacy risks remain. Differential privacy, as in option C, is the most appropriate control to ensure that outputs do not reveal whether particular personal data was used.

ISACA, AAIA Exam Content Outline -- Domain 1: Privacy and Data Governance Programs; Domain 2: Data Management Specific to AI (data confidentiality, data security).

ISACA guidance on privacy-by-design and AI risk management concepts reflected in AAIA.