Corella: A Private Multi Server Learning Approach based on Correlated Queries
This addresses privacy concerns for mobile device users offloading ML tasks to cloud/edge servers, offering an incremental improvement over existing methods like noise addition, MPC, and HE.
The paper tackles the problem of preserving client data privacy in cloud/edge machine learning by proposing Corella, a multi-server approach that adds strong, correlated noise to data to prevent information leakage while enabling accurate model outputs through server output combination. Simulation results show high accuracy for classification and autoencoder tasks.
The emerging applications of machine learning algorithms on mobile devices motivate us to offload the computation tasks of training a model or deploying a trained one to the cloud or at the edge of the network. One of the major challenges in this setup is to guarantee the privacy of the client data. Various methods have been proposed to protect privacy in the literature. Those include (i) adding noise to the client data, which reduces the accuracy of the result, (ii) using secure multiparty computation (MPC), which requires significant communication among the computing nodes or with the client, (iii) relying on homomorphic encryption (HE) methods, which significantly increases computation load at the servers. In this paper, we propose $\textit{Corella}$ as an alternative approach to protect the privacy of data. The proposed scheme relies on a cluster of servers, where at most $T \in \mathbb{N}$ of them may collude, each running a learning model (e.g., a deep neural network). Each server is fed with the client data, added with $\textit{strong}$ noise, independent from user data. The variance of the noise is set to be large enough to make the information leakage to any subset of up to $T$ servers information-theoretically negligible. On the other hand, the added noises for different servers are $\textit{correlated}$. This correlation among the queries allows the parameters of the models running on different servers to be $\textit{trained}$ such that the client can mitigate the contribution of the noises by combining the outputs of the servers, and recover the final result with high accuracy and with a minor computational effort. Simulation results for various datasets demonstrate the accuracy of the proposed approach for the classification, using deep neural networks, and the autoencoder, as supervised and unsupervised learning tasks, respectively.