DP-ACO: Differentially Private Average Consensus Optimization in Decentralized Learning

This paper introduces DP-ACO, a novel system that enables differentially private average consensus optimization within a decentralized learning framework. DP-ACO enables collaborative training of a network comprised of n nodes on aggregated data while protecting the privacy of each node’s local information. We evaluate DP-ACO using a reasonable privacy budget of \(\epsilon < 10\) , contrasting its performance with a baseline system employing Stochastic Gradient Descent (SGD) without privacy ( \(\epsilon = +\infty \) ). Our results reveal that DP-ACO and the baseline system exhibit similar utility levels, particularly in testing accuracies. The results show the effectiveness of DP-ACO in reaching the intended privacy goals while simultaneously maintaining the overall predictive capability and accuracy of trained models. DP-ACO offers a path for collaborative training on sensitive data sources in industries such as healthcare, finance, and other fields where data privacy is essential.