Poisoning Attack for Inter-agent Transfer Learning

Zelei Cheng, Zuotian Li

Abstract: In reinforcement learning, high sample complexity is a big challenge to deal with. Inter-agent transfer learning is one solution to this challenge that can leverage the experience of other more competent agents. In this paradigm, a student can make a query to the teacher and the teacher will give some action advice given the current state. However, most previous works ignored the instruction reliability problem. In this work, we investigate the instruction reliability issue based on the one-to-one teaching framework and formulate the poisoning attack as an optimization problem. By solving the optimization problem, the attacker can significantly influence the performance of the student in three different query models. Evaluation highlights that we need to consider the instruction reliability when using teacher-student frameworks in reinforcement learning.