Task 009: Multi-modal distributed learning

Abstract:

Reinforcement learning, like any other ML algorithms, poses a threat from adversarial manipulations. Adversaries can use a variety of attack models to manipulate the model either in the training or the inference phase leading to decreased accuracy or poor policies. This talk focuses on adversarial attacks on Multi-task federated Reinforcement Learning (MT-FedRL) where multiple RL agents in multiple environments try to jointly maximize the sum of individual discounted returns by sharing their policy parameters. We argue that the general adversarial methods are not good enough to create an effective attack on MT-FedRL and propose a model-poisoning attack methodology AdAMInG based on minimizing the information gain during the MT-FedRL training. We address the adversarial attack issue by proposing a modification to the general FedRL algorithm, ComA-FedRL, that works equally well with and without the adversaries. We show the effectiveness of the proposed methodology on small and medium-size problems of GridWorld and drome autonomous navigation.

 

Speaker’s Bio: 

Aqeel Anwar is an ECE Doctoral candidate at Georgia Institute of Technology working with Dr. Arijit Raychowdhury towards enabling edge intelligence in resource-constrained autonomous systems for Reinforcement Learning (RL) based applications. He completed his bachelor’s degree in Electrical Engineering from University of Engineering and Technology (UET) Lahore Pakistan and master’s degree in Electrical and Computer Engineering from Georgia Tech in 2017. Before beginning his Ph.D., Aqeel worked as a Machine Learning Engineer for a German-based startup on self-driving cars. His other interests include 3D paper modeling, graphic designing, and writing articles about machine learning and Ph.D. life in general.