Resilient Distributed Machine Learning in Secure Navy Tactical Networks
Abstract
Project AbstractApproved for Public ReleaseThis project focuses on developing frameworks and algorithms for secure, distributed,andonline machine learning in tactical networks that are characterized by agent heterogeneityand autonomy, constraints on computational and energy resources, and stringent performance requirements,particularly on latency. This new paradigm of distributed machine learning, based onon-device computing and device-to-device communication at the edge of the network, has receivedsignificant attention lately. However, critical issues of security and resilience have been relativelyless explored and the developments have mostly been fragmented. This is primarily due to the factthat distributed ML has failed to account for significant security and resiliency challenges of tacticalnetworks. For example, distributed ML has not addressed the need to detect unattended-agentcapture, followed by agent removal and replacement without human intervention beyond simplerelease of new agents; nor has it addressed the need to assure both private and authentic inter-agentcommunication despite traffic-analysis and message-injection attacks. Also, relying on a commonIoT infrastructure, distributed ML has inadvertently adopted Internet protocols with large attacksurfaces, and has not exploited the geo-proximity of ML agents to create new secure and robustprotocols. That is, it has inherited all Internet vulnerabilities caused by dependencies on DNS,PKI, NTP at the network edge simply because research attention focused exclusively on the newalgorithmic paradigms. This project aims to fill this gap by developing principled approaches forsecure and resilient online machine learning in distributed agent networks. Specifically, by employingscalable randomized key-predistribution-based mechanisms in conjunction with resilient distributedalgorithms that exploit task specific redundancies, we propose to develop methodologiesand an integrated framework for online distributed machine learning with performance guaranteesand security assurances against broad classes of adversarial attacks and anomalies (including nodecapture and data injection attacks).With the Navy#s strategic focus on Distributed Maritime Operations, distributed algorithmicprocessing at the edge and embedded on platform becomes increasingly important to integrategeographicallydistributed naval forces and synchronize operations across all domains. The outputof this research would directly inform the development of and enhanced security for the NavalOperational Architecture as the connective tissue between sensors, platforms and decision makersin adversarial and contested environments - enabling faster decisions, leveraging information fromacross the fleet, integrating manned and unmanned assets and more. This project will leverage themilitary experience and knowledge of Navy personnel and ROTC Midshipmen in the Pittsburghregion leveraging the Hacking for Defense methodology to explore theproblem/research approachand to develop a high-level Concept of Operations that will be used for a technical demonstrationof secure and distributed machine learning prototype algorithms and capability.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 03, 2023
- Source ID
- N000142312275
Entities
People
- Osman Yagan
Organizations
- Carnegie Mellon University
- Office of Naval Research
- United States Navy