Adversarial Exploitation of Policy Imitation

This paper investigates a class of attacks targeting Typically, the settings of imitation learning are concerned the confidentiality aspect of security in Deep with learning from human demonstrations. However, it is Reinforcement Learning (DRL) policies. Recent straightforward to deduce that the techniques developed for research have established the vulnerability of supervised those settings may also be applied to learning from artificial machine learning models (e.g., classifiers) experts, such as DRL agents. Of particular relevance to to model extraction attacks. Such attacks leverage this research is the emerging area of Reinforcement Learning the loosely-restricted ability of the attacker to iteratively with Expert Demonstrations (RLED)[Piot et al., 2014]. The query the model for labels, thereby allowing techniques of RLED aim to minimize the effect of modeling for the forging of a labeled dataset which can be imperfections on the efficacy of the final RL policy, while used to train a replica of the original model. In this minimizing the cost of training by leveraging the information work, we demonstrate the feasibility of exploiting available demonstrations to reduce the search space of imitation learning techniques in launching model the policy.