Distributed Intelligence Optimization under Interference in Heterogeneous Resource-Constrained Wirel
Abstract
Approved for Public Release There is growing interest in leveraging machine learning (ML) techniques to provide advanced intelligenc,e in wireless systems. Current techniques for distributing model training over systems of devices, such as the recently popularized,federated learning approach, often assume a star topology learning architecture, which requires frequent device-to-server communic,ations. This architecture encounters fundamental challenges under three dimensions of wireless heterogeneity: (i) computation, due t,o different device processing capabilities; (ii) communication, due to varying proximities and interference conditions; and (iii) st,atistical, due to varying data collection environments across devices. As the complexity of ML models continues to grow, these heter,ocess. The incurred overhead becomes particularly problematic in settings such as electronic warfare (EW) where size, weight, and po,wer (SWaP) reduction is an important objective that must be coupled with the needs for accurate situational awareness and rapid resp,onse to enemy threats.The objective of this project is to establish the foundation for distributed learning architectures that addre,ss the heterogeneity properties and resource constraints of congested and contested wireless systems. Proposed investigations are di,vided into three thrusts. First, a novel learning architecture consisting of interactions between local model updates (at devices),,local intelligence synchronization (among clusters of devices), and global aggregations (by the server) will be investigated, exploi,ting the proliferation of direct device-to-device (D2D) communications. Second, techniques for exploiting spatial and temporal redun,dancies that emerge across wireless systems will be developed, consisting of novel techniques for recycling model updates and intell,igently sampling devices that lead to more efficient device-to-server interactions in distributed ML. Third, consideration will be g,iven for how the availability of a few mobile, processing-capable nodes can facilitate agile intelligence management across a large-,scale wireless deployment, through joint trajectory planning and learning parameter design. These thrusts will lead to discovery of,the underlying relationships between wireless connectivity, device characteristics, and intelligence performance in the presence of,intentional and unintentional interference. These relationships will be leveraged in the orchestration of distributed learning to en,hance the achievable tradeoff between intelligence quality and resource efficiency.This project will also contain a significant expe,rimental component, through a testbed of software defined radio (SDR) devices emulating practical wireless intelligence systems with, realistic interferer nodes. Measurements collected from these experiments will be used to validate and refine the analytical result,s. We envision that our efforts will also be guided by databases of emerging wireless signals, from programs such as DARPA RF Machin,e Learning Systems (RFMLS) and any related ones being pursued by ONR. DoD advisors on, the methodologies, testbed construction, and interferer models considered.If successful, the proposed research will lead to robust,,gence, Surveillance, and Reconnaissance (ISR) missions carried out by a heterogeneous set of land, sea, and air assets. Rapid D2D lo,cal synchronizations among ISR assets coupled with improved uplink communication efficiency to access points will enable scalable tr,aining and execution of ML models across systems that identify and respond to adversarial threats.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 01, 2022
- Source ID
- N000142212305
Entities
People
- Christopher Brinton
Organizations
- Office of Naval Research
- Purdue University
- United States Navy