Research Team Status
- Names of researchers and position
(e.g. Research Scientist, PostDoc, Student (Undergrad/Masters/PhD))- Sandeep Neema – PI
- Gabor Karsai – co-PI
- Ankita Samaddar- Postdoctoral Scholar
- Robert Canady - Postdoctoral Scholar
- Nicholas Potteiger – PhD student
- Noah Dahle - PhD student
PI Koutsoukos accepted a Program Manager position at DARPA I2O starting in June. Dr. Sandeep Neema, Professor in the Department of Computer Science and Director of the Institute for Software Integrated Systems at Vanderbilt University assumed the role of the PI for the award (with the help of Gabor Karsai) after receiving approval from the NSA SoS Program Directors. Dr. Neema is an expert in Neurosymbolic AI, and is working closely with the research group. He has been leading multiple synergistic efforts. We are confident that this transition will ensure continuity and maintain the high quality of research under this award.
- Any new collaborations with other universities/researchers?
- Collaboration with the DARPA CASTLE project at Vanderbilt and University of Virginia for evaluation of the neurosymbolic cyber-agents using a realistic emulation testbed.
Project Goals
- What is the current project goal?
- Develop advanced learning methods for integrating learning-enabled components to behavior trees.
- Design robust distributed cyber-defense agents using evolving behavior trees (EBTs) for CAGE Challenge 3 and 4.
- Develop neuro-symbolic representations for red agents using advanced learning methods.
- Evaluate the EBT-based agents in computer network defense scenarios based in the CybORG simulation environment and in the Vanderbilt emulation testbed developed under the DARPA CASTLE.
- How does the current goal factor into the long-term goal of the project?
- The current goals address the development of the advanced learning methods and evaluation which are the main tasks of the option period 1.
Accomplishments
- Address whether project milestones were met. If milestones were not met, explain why, and what are the next steps.
- Project milestones are met with respect to both the agent architecture, advance learning methods, and the demonstration/evaluation. In summary, we extend our approach from base year 1 to incorporate advanced learning methods, design distributed cyber agents, and evaluation of the techniques using the CybORG simulator. We developed new approaches to predict the policies being used by red agents at runtime. The predicted red policies can be used during runtime to guide the actions of the blue agent. We continue to evaluate the developed methods in the CybOrg simulator and the Vanderbilt emulation testbed developed under the DARPA CASTLE program.
- Project milestones are met with respect to both the agent architecture, advance learning methods, and the demonstration/evaluation. In summary, we extend our approach from base year 1 to incorporate advanced learning methods, design distributed cyber agents, and evaluation of the techniques using the CybORG simulator. We developed new approaches to predict the policies being used by red agents at runtime. The predicted red policies can be used during runtime to guide the actions of the blue agent. We continue to evaluate the developed methods in the CybOrg simulator and the Vanderbilt emulation testbed developed under the DARPA CASTLE program.
- What is the contribution to foundational cybersecurity research? Was there something discovered or confirmed?
- Our preliminary results demonstrate that the advanced learning methods for neurosymoblic agents can be used to detect to improve robustness and generalizability.
- Our preliminary results demonstrate that the advanced learning methods for neurosymoblic agents can be used to detect to improve robustness and generalizability.
- Impact of research
- Internal to the university (coursework/curriculum)
- External to the university (transition to industry/government (local/federal); patents, start-ups, software, etc.)
- Any acknowledgements, awards, or references in media?
- A paper from the project team was accepted to NeuS 2025: The 2nd International Conference on Neuro-symbolic Systems, held May 28–30, 2025, at the University of Pennsylvania. The paper, titled "Real-Time Reachability for Neurosymbolic Reinforcement Learning based Safe Autonomous Navigation", was authored by Nicholas Potteiger, Diego Manzanas-Lopez, Taylor T. Johnson, and Xenofon Koutsoukos. The work presents a novel method for integrating real-time reachability analysis with neurosymbolic reinforcement learning to enhance safety and efficiency in autonomous navigation. The team developed a software package, RusTReach, and demonstrated its performance on an embedded platform in a quadcopter navigation task. The approach outperformed existing methods while ensuring safety and real-time compliance. Code and demonstration videos are publicly available at https://github.com/npotteig/rustreach.
Publications and presentations
- Add publication reference in the publications section below. An authors copy or final should be added in the report file(s) section. This is for NSA's review only.
- Optionally, upload technical presentation slides that may go into greater detail. For NSA's review only.