Biblio

The appearance of distributed learning allows multiple participants to collaboratively train a global model, where instead of directly releasing their private training data with the server, participants iteratively share their local model updates (parameters) with the server. However, recent attacks demonstrate that sharing local model updates is not sufficient to provide reasonable privacy guarantees, as local model updates may result in significant privacy leakage about local training data of participants. To address this issue, in this paper, we present an alternative approach that combines distributed differential privacy (DDP) with a three-layer encryption protocol to achieve a better privacy-utility tradeoff than the existing DP-based approaches. An unbiased encoding algorithm is proposed to cope with floating-point values, while largely reducing mean squared error due to rounding. Our approach dispenses with the need for any trusted server, and enables each party to add less noise to achieve the same privacy and similar utility guarantees as that of the centralized differential privacy. Preliminary analysis and performance evaluation confirm the effectiveness of our approach, which achieves significantly higher accuracy than that of local differential privacy approach, and comparable accuracy to the centralized differential privacy approach.

Alternating direction method of multiplier (ADMM) is a powerful method to solve decentralized convex optimization problems. In distributed settings, each node performs computation with its local data and the local results are exchanged among neighboring nodes in an iterative fashion. During this iterative process the leakage of data privacy arises and can accumulate significantly over many iterations, making it difficult to balance the privacy-utility tradeoff. In this study we propose Recycled ADMM (R-ADMM), where a linear approximation is applied to every even iteration, its solution directly calculated using only results from the previous, odd iteration. It turns out that under such a scheme, half of the updates incur no privacy loss and require much less computation compared to the conventional ADMM. We obtain a sufficient condition for the convergence of R-ADMM and provide the privacy analysis based on objective perturbation.