Building Secure, Resilient, and Connected Systems
The explosion of data resulting from the propagation and interconnection of sensors and systems over the last decade has created opportunities for enhancing the impact of AI and autonomous agents. This thrust will include data-driven discovery to address fundamental questions at the frontiers of science and engineering. The research in this thrust will result in the convergence of expertise among researchers including domain scientists, data scientists, and cyberinfrastructure experts to form novel frameworks, synergies, and concepts to implement the vision and to frame data-analytic challenges in new ways in science and engineering in general and those that directly impact the operations of the armed forces in particular. Included in this thrust is research related to networking and sensing, edge intelligence, health assessment, and cross-domain learning. In addition, the ubiquity of cyber-physical systems has also created more opportunities for adversaries to exploit, disrupt or poison them to gain information advantages or cause disruption or harm. Harnessing data is essential to enabling and extending a virtual proving ground. Many of the research activities below address dual-use IoT issues that extend well beyond military applications.
Resilient and Effective Communication and Sensing for Autonomous Vehicles and Systems
This project seeks to realize the full potential of autonomous vehicles and systems through reliable connectivity, efficient communication, and resilient sensing in contested multi-domain environments. Anticipated Outcomes: the design and selection of interoperable sensing, communication, and networking protocols; an ontology that can reason over assets, networks, communication protocols, communication hardware, and mission context to dynamically infer strategy that maintains communication and mission objectives in the presence of attacks and failures; simulation and experimental testbeds to measure the efficiency of sensing and communicating data in a closed or open environment; assessment of communication performance in non-line-of-sight scenarios with moving robotic platforms; quality and confidence metrics for optimal sensor placement; self- and semi-supervised semantic segmentation techniques for fine-grained terrain analysis; and an evaluation of terrain analysis techniques using public datasets in rugged environments.
Collaborative AI Models for Real Time Situation Awareness
The project seeks to develop and demonstrate new algorithms and system-level optimizations for situational understanding on collaborative distributed platforms. Anticipated Outcomes: reinforcement learning-based approaches to sensor placement that are optimal for collaborative perception and situational awareness; a comparison of designs that maximize coverage vs. maximizing fidelity in poor visibility regions; a fusion framework for combining semantic segmentation with camouflaged/obfuscated object detection for wide area surveillance; coordination strategies for multiple autonomous UGVs to collaboratively perform deep ground surveillance; characterization of performance bottlenecks in sensing, communication and coordination among heterogeneous agent; a hierarchical episodic memory framework to store state-macro action pairs and improve the efficiency of episodic memory; experimental demonstration of the proposed framework on real world environments using ARL-compatible platforms and software.
Cyber Defense for Unmanned Vehicles and Autonomous Systems
This project will develop algorithms and methodologies for protecting against data and model poisoning attacks, securing autonomous vehicle systems including unmanned aerial (UAV) and ground vehicles (UGV) and the corresponding communication network that enables autonomous systems to work in tandem rather than in isolation. Anticipated Outcomes: demonstration of how inference attacks can be launched and prevented on UGV-mounted AI platforms, how unmanned assets and soldier-borne communication devices can be detected using encrypted network traffic data, and how army assets can be attacked using electronic control units and how such attacks can be prevented.
Cross-Domain Learning with Few Labels
This project seeks to enable fast and efficient human-agent teaming by understanding contested environments that require aggregation of information from a multitude of data sources and responding to the situation. Anticipated Outcomes: an extension of the existing knowledge graph and underlying ontology to represent a selected scenario, integrate knowledge representation with other subtasks, and evaluate querying of the knowledge graph; collection of multiview data using cameras on the ground and in the air; use of partially labeled data to build view-invariant models; validation of the proposed model using public datasets; multi-source domain adaptation-based semantic segmentation fusing camera, thermal, and lidar sensor data for semantic segmentation and transfer learning; and mechanisms by which robots can learn from and communicate using spoken language, particularly as it pertains to anomalous and/or atypical situations.
Re-Deployable and Scalable Automatic Target Recognition via Domain Invariant Agglomerative Clustering
This project will generalize the application of multi-domain automatic target recognition assets efficiently and effectively across various operational domains. Anticipated Outcomes: unsupervised automatic target recognition algorithms; target re-identification algorithms co-designed with dataflow methodologies to systematically balance system constraints; integration and demonstration of automatic target recognition algorithms using the ARL aerial autonomy stack.
Scarce Data Machine Learning for Energetic Materials Research
This project will develop machine-learning (ML) approaches suitable for scarce-data problems for the study, design, and analysis of the properties of energetic materials (EM). Anticipated Outcomes: featurization of chemical strains for prediction of energetic performance; featurization and data fusion techniques for ML study of ionic polarizable energetic materials; and approaches for combining/connecting/relating chemical space representations with technical language embeddings.
Slicing-Enabled Private 5G Network to Support AI and Autonomy for Multi-Agent Systems
This project seeks to design, analyze, and evaluate private 5G radio resource management to enable artificial intelligence and autonomy for multi-agent systems in mission-critical applications by systematically addressing radio resource management issues. Anticipated Outcomes: the wireless connectivity requirements to ensure a prescribed level of performance for various application scenarios; an intelligent radio resource slicing strategy to create slices for different applications and devices; and a testbed to demonstrate the end-to-end performance of the proposed slicing mechanism.
Location Deception in Mobile Ad Hoc Network
This project seeks to develop a theoretical foundation and a set of practical solutions for connected soldiers, vehicles, drones, and/or sensors to securely locate and share their real-time position. Anticipated Outcomes: a problem formulation and threat models for scenarios of interest; evaluation of existing solutions and identification of research needs and challenges; new position locating mechanisms and assessment; secure communication protocol under the threat model and preliminary assessment; necessary/sufficient conditions for the problem to be solvable with representative examples; and one baseline and one sophisticated deception protocol with preliminary simulation results.