Provably Robust Dynamic Systems
Abstract
Dynamic systems based on deep neural networks have many DoD-relevant applications ranging from robotics and autonomous planning to multi-modal, context-adapted data fusion. Dynamic models, however, can be quite sensitive to small adversarial perturbations of their inputs where adversaries can induce time-dependent perturbations at multiple scales to change the outputs of the systems in specific times or more broadly. Such adversarial perturbations can be small and thus imperceptible to humans, making their security risks even more worrisome in highly sensitive applications where model reliability and trustworthiness are critical. Most of the existing literature on adversarial robustness study empirical robustness of static models such as image classification problems. Such techniques, however, fall short against strong and adaptive adversarial attacks that are particularly designed to break dynamic learning systems because in dynamic settings, the adversary can adapt its strategy to the defense applied by the victim in previous time steps. In this proposal, we aim to resolve these issues by providing a comprehensive and fundamental understanding of provable robustness in dynamic and adaptive learning, leading to practically useful methods with theoretical guarantees. Towards this goal, we will leverage tools and concepts from optimization, deep learning, machine learning, information theory and statistics to (i) develop provably robust dynamic sequential decision making methods, (ii) study dynamic learning in zero-shot avoidance, (iii) investigate provable robustness in multi-agent reinforcement learning problems and (iv) develop provably robust methods against online adversarial attacks. Our proposed efforts will shed light on some fundamental issues of robust dynamic learning and will lead to reliable and practical methods with provable performance guarantees.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 27, 2023
- Source ID
- W911NF2310297
Entities
People
- Soheil Feizi
Organizations
- Army Contracting Command
- United States Army
- University of Maryland