Designing distributed learning algorithms that are both efficient and private

Develop distributed learning algorithms for decentralized settings that simultaneously achieve efficiency in computation and communication while providing rigorous differential privacy guarantees for participants’ data.

Background

The paper studies distributed learning over multi-agent networks, where agents collaboratively optimize a global objective using local datasets while communicating over an undirected graph. Two core design goals are emphasized: computation-efficiency (e.g., via stochastic gradients) and communication-efficiency (e.g., via local training to reduce communication frequency).

Differential privacy is required to protect agents’ sensitive data from being inferred through shared models. Although there has been extensive work on differentially private distributed learning, the authors explicitly state that achieving both efficiency and privacy in a single algorithm remains an open challenge, motivating their proposed LT-ADMM-DP approach.