CIF: Small: Ensuring Accuracy in Differentially Private Decentralized Optimization

Clemson University, Clemson SC

Investigators

Abstract

Advances in wireless communications and low-cost computing devices have enabled a proliferation of large distributed networks of data collection systems, constituting a major component of the emergent Internet of Things (IoT), Intelligent Transportation Systems (ITS), and the Smart Grid (SG) . Complementing these advances is the significant progress in decentralized optimization software that enables the basic functionalities of such distributed networked systems, including cooperative control, network information fusion, network coordination, and distributed data mining/learning. However, information sharing over such large networks creates vulnerabilities and concerns about privacy, which can be especially acute in privacy-sensitive applications such as smart metering and connected vehicle networks. Differential privacy is the most widely used protective mechanism for privacy due to its simplicity, scalability, and strong resilience against attempts to recover sensitive information from post-processed data. However, all existing differential-privacy solutions for decentralized optimization face the dilemma of how to achieve data privacy protection by compromising the optimizer's speed of convergence rather than its accuracy. This project leverages on the PI’s recent discovery that it is possible to achieve differential privacy guarantees without compromising utility by leveraging on the optimizer's speed of convergence rather than its accuracy. Specifically, the project will establish theoretical and algorithmic foundations for the problem of ensuring differential privacy in decentralized optimization without losing provable optimality. In addition to broadly enabling more effective privacy protections for decentralized networks, the project will impact education by enriching the current curriculum on control and networked systems, and training undergraduate and graduate students in interdisciplinary information privacy research and its applications. This project will establish theoretical and algorithmic foundations for ensuring differential privacy in decentralized optimization algorithms without losing provable optimality. The main research thrusts are to: (1) Investigate the sacrifice in convergence speed in differentially private decentralized optimization with provable optimality using an information-theoretic approach; (2) Explore and establish differential privacy without losing provable optimality in decentralized online optimization, where data are not pre-collected before implementing the algorithm but rather are acquired in a sequential manner; (3) Explore and establish differential privacy without losing probable optimality in decentralized optimization algorithms subject to shared coupling constraints among participating agents’ decision variables; (4) Explore and establish differential privacy in decentralized Nash games (which are essentially decentralized optimization problems with noncooperative agents) without losing provable optimality; (5) Evaluate obtained results using numerical simulations as well as experiments on real-word distributed systems in smart grids and networked intelligent vehicles. This project is jointly funded by Core Program of the Computing and Communication Foundations Division (CCF) and the Established Program to Stimulate Competitive Research (EPSCoR). This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.