Learning and Computational models for Multi-Domain Operations in Tactical Edge Networks
Abstract
Motivation: Battlefield operations increasingly rely on IoT sensors and unmanned systems interacting with military personnel to carry out missions, and relatively powerful edge devices to preprocess and forward them to the military cloud. Mining the records that document the events unfolding can provide predictive models that streamline human-machines teams interactions and help leverage optimally the concurrent multi-modal sensor information. To this end, a cohesive mathematical representation for the data is needed. New algebraic models are emerging for multi-domain data that rely on the notion of graph signals; these algebraic models have the potential to become a powerful alternative to Bayesian graphical models. Unlike images and sequential data, in order to extract informative features from multi-domain graph signals, one needs to learn the underlying latent graph representation that explains the events recorded in the Army network, as well as in the enemyÕs network. Another trend is shifting computations at the edge devices, to reduce the latency of offloading them to the military cloud. While great advances have been made in using locally machine intelligence, using it across multi-domain network operations and harnessing edge devices to train the machine learning algorithms, requires further studies, to handle data that need to be pieced together from multiple sources as well as adversarial attacks. Technical contribution: This project focuses on filling the aforementioned gaps: 1) the gap in learning networked system interaction models, where data come from multiple domains: manned and unmanned machine automata, humans, and sensors, exposed to coupled attack vectors and 2) the gap in solving learning tasks associated to multi-domain graph signals through a prototypical military network of multiaccess edge computation (MEC) resources, operating in a challenging adversarial environment. The research is divided in two main thrusts. The first thrust advances graph-based learning models and methods to capture the dynamic environment and multi-domain network interactions and sensing occurring at the military tactical edge. The second thrust focuses on methods to harness tactical multiaccess edge computing (MEC) resources, by decomposing federated learning tasks so that they can work in a dynamic, robust meshed configuration that more faithfully reflects military networks challenging environment. The first thrust is motivated by the observation that machine and human interactions, the dynamic sensor information that influences their actions, and the various vectors of attacks they are exposed to can be viewed as multimodal signals over an irregular support, i.e. a graph, with blue-team and red-team nodes. The research focuses on the problem unveiling the latent graph structure that explains the observations, leveraging insights from graph signal processing (GSP) and extending their applicability to multimodal, dynamic interactions. The second thrust is aimed at leveraging MEC resources to replace the paradigm of offloading computations to a remote military cloud. We complement our modeling analysis with self-healing and hierarchical edge computing algorithms that can achieve optimality tuning on the challenging network environment of the tactical edge. The latter will support the tracking of network interactions, following the models introduced in the first thrusts. Benefits to army operations: The project technical contributions will improve the capability of the military tactical edge to see, isolate and converge in its multi-modal operations to defeat the enemy. In its first thrust it will help modeling the interactions that occur at its edge as well as those that occur in the enemy network and formulate the learning. In its second thrust it will provide novel algorithms that can cope with the hierarchical dynamic
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 25, 2022
- Source ID
- W911NF2210228
Entities
People
- Anna Scaglione
Organizations
- Army Contracting Command
- Cornell University
- United States Army