Privacy-Utility Tradeoff

This field asks how much sensitive information can be protected without undermining tasks such as medical prediction, recommendation, speech analysis, and agent workflows. Approaches include differential privacy in central, local, and shuffled settings; federated and synthetic data methods; anonymization and representation learning; and selective disclosure that reveals only task-relevant details. Evaluations compare task performance with risks such as membership inference, re-identification, and attribute inference, sometimes also measuring compute costs and fairness. A central tension is that formal protections can substantially reduce utility, while task-specific safeguards may preserve it but depend on realistic threat models and consistent evaluation.