DPLE: A Privacy-Enhanced and Straggler-Resilient Distributed Learning Framework for Smart Cloud

Distributed machine learning (DML) encounters issues related to privacy and the presence of straggling nodes in smart cloud systems. Lagrange coded computing offers a partial solution to mitigate these concerns. Nonetheless, the privacy of the system becomes vulnerable when the number of semi-trusted nodes surpasses a specific limit, or when external eavesdroppers are present. To confront this hurdle, we introduce a novel framework for distributed learning called DPLE (Differentially Private Lagrange Encoding). This framework employs Lagrange interpolation polynomials to obscure the original data while introducing redundancy, thus improving privacy safeguards and increasing robustness to straggling nodes. It also incorporates artificial noise into local computation outcomes to protect confidential data from potential exposures. Furthermore, we perform theoretical analyses to identify the necessary variance of this noise to maintain desired privacy levels. Experimental validations confirm the efficacy of DPLE and examine how different settings of system parameters impact the accuracies of the results.