Multi-Cloud Privacy Preserving Schemes for Linear Data Mining

This paper presents an approach to privacy-preserving data mining that relies upon a relatively simple secret sharing scheme. Its main feature is that users, sensitive data owners, are engaged in the secret sharing operations that protect their privacy. They are grouped in independent clouds connected to a central unit, the data miner, that only manages the aggregated data of each cloud, therefore avoiding the disclosure of information belonging to single nodes. We propose two privacy preserving schemes, with different privacy levels and communication costs. When designing them, we assume that some users' data might become inaccessible during the operation of the privacy preserving protocols, due to intermittent network connectivity or sudden user departures, and therefore introduce a new performance metric, the failure probability, defined as the probability that the mining output cannot guarantee the desired level of accuracy. We then discuss the attractive trade-offs between privacy, accuracy and communication overhead that each scheme exhibits.