2021/02/18 by Parham Gohari, Gohari, Parham, Bo Chen +7
Computer Science · #Age of Information Optimization #Artificial Intelligence (cs.AI) #Cryptography and Security (cs.CR) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Mobile Crowdsensing and Crowdsourcing #Privacy-Preserving Technologies in Data
paper · pdf · doi:10.48550/arxiv.2102.09599
openalex publication_date 2021/02/18 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Kickstarting deep reinforcement learning algorithms facilitate a\nteacher-student relationship among the agents and allow for a well-performing\nteacher to share demonstrations with a student to expedite the student's\ntraining. However, despite the known benefits, the demonstrations may contain\nsensitive information about the teacher's training data and existing\nkickstarting methods do not take any measures to protect it. Therefore, we use\nthe framework of differential privacy to develop a mechanism that securely\nshares the teacher's demonstrations with the student. The mechanism allows for\nthe teacher to decide upon the accuracy of its demonstrations with respect to\nthe privacy budget that it consumes, thereby granting the teacher full control\nover its data privacy. We then develop a kickstarted deep reinforcement\nlearning algorithm for the student that is privacy-aware because we calibrate\nits objective with the parameters of the teacher's privacy mechanism. The\nprivacy-aware design of the algorithm makes it possible to kickstart the\nstudent's learning despite the perturbations induced by the privacy mechanism.\nFrom numerical experiments, we highlight three empirical results: (i) the\nalgorithm succeeds in expediting the student's learning, (ii) the student\nconverges to a performance level that was not possible without the\ndemonstrations, and (iii) the student maintains its enhanced performance even\nafter the teacher stops sharing useful demonstrations due to its privacy budget\nconstraints.\n