My research focuses on Trustworthy RL, with emphasis on policy robustness and policy privacy:

(1) We address robust RL by formulating it as a max-expectation optimization problem and introducing the DRRL framework, which dynamically generates and sequences tasks of varying difficulty to enhance policy robustness and training stability. (Li et al, IJCAI'23)

(2) We treat robust RL as a multi-task problem, using a GAN-based task generation model and curricular learning to iteratively train policies with progressive tasks, achieving improved stability and robustness across diverse environments. (Li et al, TST'22 and Tian et al, AutoSec'21)

(3) To protect the privacy of RL models from imitation learning attacks, we propose a hierarchical Policy Confusion Defense (PCD) framework that models privacy-preserving training as finding optimal dynamic switching among diverse expert policies to induce confusion. (Li et al, submitted to TDSC'23)

We target the data poisoning congestion attacks in USDOT-sponsored Intelligent Traffic Signal System:

(1) To collaborative defend against data poinsing attacks in multi-intersection traffic network, we model multi-intersection signal planning as a multi-agent RL problem and propose an actor-attention-critic algorithm to enhance robustness while improving traffic and energy efficiency. (Li et al, TGCN'22)

(2) To bridge the gap in timely visualized attack predictions, we introduce a CycleGAN-based framework leveraging traffic image feature learning to predict congestion caused by attacks. (Li et al, WCMC'20)