Nonlinear Information Fusion and its Propagation Under Uncertainty in Cyber and Cyber-Physical Domains: Algorithms and Applications

Abstract

While there has been great progress in the last ten years on data inference models using ``big data", there is much less success in formulating inference algorithms using ``small data . In fact, unlike commercial applications, in many DoD critical applications the data is very sparse and dynamic while the fidelity of the data is variable, often corrupted by adversarial data, which is blended or disguised together with the useful data. Based on this information and corresponding decisions, other operations may be required to be performed downstream in another domain of cyber or physical nature. Currently, there are no scientific methods to model computationally this propagation of information as existing data inference techniques even for streaming data are relying on high-fidelity, the so-called ``gold data". Moreover, there are no rigorous mathematical formulations that track the uncertainty in the information fusion from various diverse sources of variable fidelity and its propagation across cyber and physical domains. Here we address these problems by formulating appropriate abstractions of such scenarios, and we propose to develop robust nonlinear information fusion algorithms that safeguard against erroneous and contaminated data. We also address the problem of propagating this reconstructed information through cyber or physical domains, and propose new domain decomposition algorithms that propagate not only the mean (averaged) prediction but also its uncertainty seamlessly from one domain to another. The general set up consists of some information obtained from sources we trust (albeit very few) and much more information obtained from sources of lower fidelity. Based on preliminary work, we can argue that accurate response surfaces can be obtained by deep learning techniques relying on Gaussian process regression that tracks the response and its uncertainty at every level of fidelity. However, we need to generalize such algorithms to cases with information signals containing significant frequency errors and overall trends that contradict the trends of the high-fidelity data, and we propose to develop theoretical guarantees for such cases. In addition, we will develop local reconstruction procedures tailored to specific type of information and coupled to another cyber or physical domain where such information may be propagated. This domain coupling is a natural extension of our previous ARO grant focused on the propagation of stochastic solutions across domains characterized by heterogeneous solutions with vastly different correlations lengths. In this one year project, we propose three specific tasks on (1) Robust multi-fidelity information fusion; (2) Cyber-cyber domain coupling; and (3) Cyber-physical domain coupling. We will demonstrate the new algorithms with prototype applications that we will develop jointly with researchers at the Army Research Lab in the context of multiscale modeling of materials, mesoscopic weather modeling, and cyber-physical decision making applications.

Document Details

Document Type
DoD Grant Award
Publication Date
Oct 16, 2018
Source ID
W911NF1810301

Entities

People

  • George Karniadakis

Organizations

  • Army Contracting Command
  • Brown University
  • United States Army

Tags

Fields of Study

  • Computer science

Readers

  • Distributed Systems and Data Platform Development
  • Theoretical Analysis.
  • Wave Propagation and Nonlinear Chaotic Dynamics.

Technology Areas

  • AI & ML
  • AI & ML - Bayesian Inference
  • AI & ML - Machine Learning Algorithms
  • Cyber
  • Cyber - Cryptography