Privacy Preservation

Homomorphic encryption enables computation on encrypted data without decrypting it. Differential privacy protects individual records by adding calibrated noise to query outputs. k-anonymity makes an individual's record indistinguishable from at least k-1 others. Aggregate data and training membership can still expose individuals through reconstruction and membership inference attacks. Differential privacy has been applied to federated learning and used by the US Census Bureau. Value-of-information analysis can evaluate whether data merging justifies privacy costs.