Answers are grounded in this paper and cite their sources.
In emerging distributed networks and systems, large-scale data are frequently stored across cloud, edge, and multi-center nodes, making centralized statistical analysis infeasible and creating privacy leakage risks during the transmission of gradients, parameters, and intermediate statistics. To address these issues, this paper proposes a communication-efficient differentially private parameter estimation framework for distributed statistical inference. Two privacy-preserving schemes are designed by injecting multivariate Gaussian noise into transmitted first-order gradients and into both initial parameters and gradients. The analysis considers (ε, δ)-differential privacy under calibrated noise and composition accounting. Under the stated regularity, initialization, partition-growth and noise-scale conditions, the estimators remain statistically consistent and retain the convergence order of the non-private one-step estimator. Monte Carlo simulations under ε=0.5 and δ=10−4 show that the private estimators converge to the true parameter values as the number of machines or local sample size increases, while the privacy budget controls the expected privacy-accuracy tradeoff. A real-world case study on the U.S. Airline Database further demonstrates that the relative deviations between privacy-preserving estimators and non-private baselines remain around 2%, and the per-node runtimes of the two private schemes are only 0.0118 s and 0.0123 s, compared with 19.61 s for centralized computation on the sampled dataset. These results indicate that the proposed method can provide rigorous privacy protection, high communication efficiency, and practically acceptable estimation accuracy for privacy-sensitive distributed data analytics.
No takes yet. Share an insight, caveat, or question.
Yanling Wang (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: