PhD student @ Purdue CS
I am a PhD student of Computer Science Department, Purdue University. My advisor is Suresh Jagannathan. I am working on attack/defense/assurance of deep learning, especially deep reinforcement learning, from the perspective of formal methods. I received my B.Eng. of Software Engineering from University of Electronic Science and Technology of China in 2018.
CV updated at 2021-04-07.
echo "emlrYW5neGlvbmdAZ21haWwuY29t" | base64 -d
Although deep neural networks have achieved promising performance in various tasks, they are generally used as black-box functions without any formal guarantee on their properties. For example, without formal analysis, it is unclear whether a neural-network-controlled drone, which is operated in a complex environment, will collide with the ground or not. This line of work provided verifiable safety guarantee for cyber-physical-systems (e.g., robots, UVA) trained with deep reinforcement learning.
Scalable Synthesis of Verified Controllers in Deep Reinforcement Learning
Zikang Xiong, and Suresh Jagannathan.
Under Review [pdf]
Neural network controllers are not robust to adversarial attacks, which exposes them to great threats from malicious attackers. We aim to explore both attack and defense techniques for deep-neural-network controlled systems, thus providing more robust neural network controllers.
Robustness to Adversarial Attacks in Learning-Enabled Controllers
Zikang Xiong, Joe Eappen, He Zhu and Suresh Jagannathan.
Adaptive and Learning Agents Workshop at AAMAS 2021 & Under Submission [pdf] [tool] [talk]
📚 My recent paper reading notes.