PPDOG: Privacy-Preserving Distributed Optimization With Performance Guarantees in Networked Systems
Dan Yu, Xiufang Shi, Yangfei Lin, Celimuge Wu, Jiming Chen · IEEE Transactions on Network Science and Engineering · 2025
In distributed optimization, agents collaborate by sharing decision variables or gradients during the local data processing. However, this information exchange exposes agents to privacy risks, as adversaries can potentially infer their local data. While existing privacy-preserving methods employ encryption or noise-adding mechanisms, they suffer from heavy computational overhead or degraded convergence properties, respectively. To overcome these limitations, we propose Privacy-Preserving Distributed Optimization with performance Guarantees (PPDOG), a lightweight algorithm that ensures both privacy preservation and convergence properties. PPDOG utilizes Secure Multi-party Computation (SMPC) to initially generate weighted-zero-sum perturbation noise among agents within the same local network. During the optimization process, agents securely update their perturbation noise through an efficient recursive process and add it to the exchanged parameters. These perturbations are mathematically designed to be canceled out during iterations, thereby preserving both privacy and convergence properties without introducing high computational complexity. Extensive experimental results demonstrate that PPDOG performs better than existing methods across various distributed optimization tasks by effectively balancing privacy preservation, convergence properties, and computational efficiency.