Client-Personalized Differential Privacy for Federated Subgroup Discovery in Heterogeneous Data

Abstract

Subgroup discovery returns interpretable descriptions whose outcome rates differ from a population baseline. We present FedPSD, a federated design with client-specific noise budgets and precision-weighted subgroup scoring, together with a reported simulation on UCI Adult and the scikit-learn breast-cancer dataset. Twenty eight-client federations cover ten seeds on each corpus. Personalized budgets have a size-weighted mean of 4; this matches a scalar budget-allocation convention, not each record's protection or a worst-case federation guarantee. Compared with the uniform-budget setting of 4, mean held-out top-10 WRAcc rises from 0.0346 to 0.0503 on Adult and from 0.0488 to 0.0559 on breast cancer; the reported pooled paired Wilcoxon result is W=50, p=0.040, with 16 of 20 blocks favourable. Removing precision weighting lowers Adult WRAcc to 0.0127, whereas a uniform budget of 8 under the same machinery achieves 0.0625. The count channel uses sequentially calibrated noise, but raw-data local proposals and pooled preprocessing are uncharged, so end-to-end personalized differential privacy is not established. The weighting changes the population estimand under heterogeneity. Routing and post-stratification are proposed components that were not evaluated, and the centralized search is a deeper heuristic reference rather than a proven upper bound.