Scalable Privacy-Preserving Distributed Learning

Abstract In this paper, we address the problem of privacy-preserving distributed learning and evaluation machine-learning models by analyzing it in widespread MapReduce abstraction that extend with privacy constraints. We design spindle (Scalable Privacy-preservINg Distributed LEarning), first system covers complete ML workflow enabling execution a cooperative gradient-descent obtained model preserving data confidentiality passive-adversary up to N ?1 colluding parties. uses multiparty homomorphic encryption execute parallel high-depth computations on encrypted without significant overhead. instantiate for training generalized linear datasets show is able accurately (on par non-secure centrally-trained models) efficiently (due multi-level parallelization computations) train require high number iterations large input thousands features, among hundreds providers. For instance, trains logistic-regression dataset one million samples 32 features 160 providers less than three minutes.