New research combines MPC and genetic algorithms for timely privacy-preserving optimization

A new research paper from Sebastian Gruber and colleagues tackles a core challenge in privacy-preserving distributed optimization: the runtime overhead of secure multi-party computation (MPC). In settings where multiple parties must jointly optimize a solution without revealing their private inputs (e.g., supply chain or logistics coordination), traditional MPC adds significant computational cost, often causing the optimization to miss strict deadlines. The authors propose combining MPC with evolutionary algorithms—specifically genetic algorithms and NSGA-II—to reduce the number of required MPC evaluations. The evolutionary search efficiently explores the solution space, while MPC secures each evaluation.

Experiments on the assignment problem, traveling salesperson problem, and multi-objective assignment problem demonstrate that the approach can return solutions within a deadline. A critical finding: obfuscation of evaluation results provides additional privacy against an honest-but-curious platform provider but creates a trade-off with solution quality. The paper provides empirical data on this trade-off. This work is significant for any domain where multiple entities need to collaborate on optimization under real-time constraints without leaking proprietary or sensitive data—such as finance, healthcare logistics, or federated machine learning.