<ul><li>Conformal prediction (CP) under local differential privacy (LDP) aims to preserve privacy in sensitive scenarios where the aggregator is untrusted and can only access perturbed labels.</li><li>Two approaches are proposed: in the first, users provide input features and perturbed labels using k-ary randomized response, while in the second, users add noise to their conformity score with binary search response.</li><li>Both approaches calculate the conformal threshold from noisy data without accessing true labels, ensuring data and label privacy while maintaining predictive uncertainty control.</li><li>These methods offer finite-sample coverage guarantees and demonstrate robust coverage, making them suitable for applications like medical imaging and large language model queries.</li></ul>

Privacy-Preserving Conformal Prediction Under Local Differential Privacy

Discover more