Information-theoretic privacy in federated submodel learning

We consider information-theoretic privacy in federated submodel learning, where a global server has multiple submodels. Compared to the considered conventional learning secure aggregation is adopted for ensuring privacy, provides stronger protection on selection by local machine. propose an achievable scheme that partially adopts private information retrieval (PIR) achieves minimum amount of download. With respect computation and communication overhead, we compare with naïve approach privacy.