Privacy-Preserving Alternating Direction Method of Multipliers

Abstract: As machine learning models affect our lives more strongly every day, developingmethods to train these models becomes paramount. In our paper, we focus on the problem ofminimizing a sum of functions, which lies at the heart of most - if not all - of these trainingmethods. This problem was formulated in terms of a decentralized consensus optimization, with theterms of the sum belonging to different agents. We examined the efficency and privacy-preservingproperties of methods to solve this problem, as well as conducted numerical experiments on severalvariations of the I-ADMM algorithm. Our results show that utilizing encryption is inefficientcompared to PI-ADMM1, while PI-ADMM1 converges at the same speed as I-ADMM.