Drynx: Decentralized, Secure, Verifiable System for Statistical Queries\n and Machine Learning on Distributed Datasets

Abstract

Data sharing has become of primary importance in many domains such as\nbig-data analytics, economics and medical research, but remains difficult to\nachieve when the data are sensitive. In fact, sharing personal information\nrequires individuals' unconditional consent or is often simply forbidden for\nprivacy and security reasons. In this paper, we propose Drynx, a decentralized\nsystem for privacy-conscious statistical analysis on distributed datasets.\nDrynx relies on a set of computing nodes to enable the computation of\nstatistics such as standard deviation or extrema, and the training and\nevaluation of machine-learning models on sensitive and distributed data. To\nensure data confidentiality and the privacy of the data providers, Drynx\ncombines interactive protocols, homomorphic encryption, zero-knowledge proofs\nof correctness, and differential privacy. It enables an efficient and\ndecentralized verification of the input data and of all the system's\ncomputations thus provides auditability in a strong adversarial model in which\nno entity has to be individually trusted. Drynx is highly modular, dynamic and\nparallelizable. Our evaluation shows that it enables the training of a logistic\nregression model on a dataset (12 features and 600,000 records) distributed\namong 12 data providers in less than 2 seconds. The computations are\ndistributed among 6 computing nodes, and Drynx enables the verification of the\nquery execution's correctness in less than 22 seconds.\n