NPS Junior Faculty Research Seminar

Military missions depend on critical infrastructure systems found both on our installations and outside our fencelines. This presentation overviews efforts by the Naval Postgraduate School Center for Infrastructure Defense to improve how we identify mission essential assets on our installations and in surrounding communities. Efforts within our fencelines focus on developing methods that measure worst-case disruptions across interdependent infrastructure systems and creating models that support DoD infrastructure planning and management. We focus attention on improving the DoD Mission Dependency Index (MDI) – a metric used to guide infrastructure funding across all DoD services. We identify flaws in the capacity for MDI to guide infrastructure planning decisions and provide methods to overcome them. Efforts outside our fencelines link community needs and military missions during disaster response to better prepare for future climate risks. We present case studies on infrastructure vulnerabilities that inform hazard mitigation and resilience for island and installation communities, including an analysis integrating evacuation plans for Naval Station Newport and Aquidneck Island communities to future storms. Overall, each effort centers on integrating the structure and function of infrastructure systems with the military missions they support. The goal is to provide the foundation for new methods, measures, and models for interdependent infrastructure analysis and mission assurance to compound and cascading threats.

Passive acoustic processing approaches provide utility to characterize oceanographic parameters at the relevant acoustic scales to improve awareness. For example, transiting surface vessels provide inference opportunity due to their high acoustic intensity and wide-band energy. Other localization approaches exploit the multi-path structure of a submerged source to estimate its depth and range. More recent investigations focus on 2-dimensional array-processing in the Norwegian Sea, which is of significant interest to ONR and NAVO. I will conclude this talk with an overview of my experimental, sea-going laboratory and discuss research collaboration opportunities.

Can ambient radio noise, such as the Sun and Jupiter’s radio emissions, be used for passive radar remote sensing in low-resource environments? Traditional active radars transmit a powerful electromagnetic pulse and record the echo's delay time and power to measure target properties of interest, such as range, size, velocity, and composition. Such observations are critical for early warning detection against current and evolving threats in challenging environments; however, existing radar systems are resource-intensive in terms of cost, power, mass, and spectrum usage when continuously monitoring large areas of interest. I address this challenge by presenting a novel implementation of passive radar that leverages ambient radio noise sources (instead of transmitting a powerful radio signal) as a low-resource approach for echo detection, ranging, and imaging.

This presentation covers a tailored systems engineering approach that leverages recent trends in Model-Based Systems Engineering (MBSE) that focuses on early identification of desirable system capabilities using both operational simulation models and system synthesis models. This approach is motivated by and tailored to the recent DoD emphasis on the utilization of analytic data to support decision making, which is a primary enabler of digital engineering. Digital engineering is reviewed and MBSE is demonstrated as the basis for the development and analysis of detailed system synthesis and operational models. This presentation is organized into: foundational research in both MBSE and digital engineering, definition of the application of MBSE fundamentals to support digital engineering, and a demonstration of the utility of the approach.

Deep neural networks (DNNs) are machine learning models that are being applied to increasingly diverse areas. Many of these models are used to make decisions that have substantial impact, such as those related to disease diagnosis from medical images or autonomous driving. However, can their predictions be trusted? A key component of the trustworthiness of a prediction model is proper estimation of uncertainty. Improving DNN uncertainty estimation also leads to better performance for a variety of problems, including continual learning, distributed learning, and anomaly detection. Bayesian probability tools offer a principled solution for improving DNN uncertainty estimation by learning distributions of DNNs (i.e. Bayesian DNNs). In this talk, we will briefly introduce Bayesian DNNs and demonstrate their practical usefulness using neuroimaging examples.

Ongoing technological advances will soon make defense against large numbers of adversarial drones, including swarms, a reality. With this reality comes a host of new questions related to engagement strategies, trade-off analysis, and minimum platform requirements. Answering such questions requires input from multiple scientific disciplines. In this talk, I will discuss ongoing work in our group to develop modeling, analysis, and mission-planning tools for moderate to large-scale adversarial autonomy situations. I will give special focus to our use of analysis tools from my background in the physics of glassy systems, such as many-body simulation and optimization and multivariable scaling analysis. I will show how these tools can give special insight into emergent behavior of these systems.

Computing technologies always sit within a system of people, organizations, and policies. Designing computational tools to fit this sociotechnical context is a difficult problem, often solved in ad hoc, un-validated ways. Few repeatable methods bridge reliably from policy to managerial decision-making to implementable technical requirements. We examine the characteristics and performance of assessment and governance approaches in a variety of domains, including responsible AI, privacy, cybersecurity, and system safety to understand the outlines of effective sociotechnical control for computing systems. We adapt the most effective tools and practices to evaluate if they can improve known best practice computing system governance or provide validation points for novel computing system governance approaches such as efforts to assess AI systems for legal compliance or comportment with ethical principles.

Machine learning is a trendy research topic. Its enormous popularity has encouraged plenty of great science, but also comes with several downsides. In this talk we review some of those downsides. We discuss the lack of robustness of popular methods in machine learning, a fact which severely undermines their utility in DoD applications where errors come with severe consequences. We also discuss machine learning as a tool for sowing misinformation through the creation of DeepFakes. Finally, we introduce recent NPS research efforts to combat DeepFakes by automating their detection.

Stable distributions are a versatile set of robust statistics with applications in fields far beyond network anomaly detection.  We discuss the family of stable distributions, tradeoffs in their use, and historical use cases.  We compare the stable distribution parameters to more frequently-used metrics such as Gaussian mean and the median.  We also discuss potential origins of heavy-tails / stable distributions in computer network traffic and examine the performance gains available from stable distributions and stable-derived test statistics when applied to detecting volumetric anomalies in computer network traffic.  This discussion will conclude with a quick discussion of the NPS Center for Cyber Warfare and current research projects.

Scientists and engineers increasingly make use of computational simulations in their quest to expand our knowledge of our universe and to apply that knowledge to the solution of practical problems. This year, the Department of Energy's Exascale Computing Project will, for the first time, deliver machines capable of performing a staggering 10¹⁸ arithmetic operations per second, and almost all of that computing power will come from graphics processing units (GPUs). In this talk, we will discuss the need for computing at these scales, why GPUs have been the technology chosen to take us there, and give a brief overview of some of our recent efforts developing numerical algorithms and software that can use this new hardware efficiently.

Making good decisions sequentially or over time is generally difficult, especially in uncertain (e.g., information-poor) environments. For example, which missile defense assets should be employed in the face of an incoming enemy force (e.g., missiles, aircraft), keeping in mind the possibility of future (and possibly different) requirements for those assets? One way to quantitatively model these kinds of sequential decision-making problems is as a Markov decision process (MDP).  Rooted in the study of sequential statistical decision problems during the Second World War, MDPs have since been studied in the Operations Research community with applications such as inventory control, optimal target pursuit, and vehicle routing in mind. More recently, MDPs have served as the mathematical foundation for the design of machine learning algorithms for making better decisions over time from experience in the form of numerical feedback. This line of work, known in the Computer Science and Robotics communities as Reinforcement Learning (RL), is currently a highly active research area that has seen some spectacular empirical results (e.g., AlphaGo). In this very short tutorial, we introduce MDPs and RL with an eye towards past and potential defense-oriented applications.

In the Fall of 2017, the Coupled Air Sea Processes and Electromagnetic ducting Research (CASPER) project conducted a large-scale air-sea interaction field study in the Southern California Bight. As part of a coordinated effort to characterize the marine atmospheric boundary layer from the oceanic thermal skin to the boundary layer top and inversion, the Floating Instrument Platform (FLIP) was deployed with an extensive array of atmosphere and ocean sensing systems. This expedition to Southern California would unknowingly become the iconic platform’s last science mission. Here, I’ll highlight some of the research our group did (and is still doing) using the data collected from this sunset cruise. In doing, so, I will touch on the legacy of FLIP and reflect on the continued value of stability—if not stationarity—at sea.

Low clouds rooted in the atmospheric boundary layer are inextricably connected to turbulence, and the model representation of such clouds has long been a significant source of uncertainty in weather forecasts and climate projections. In this talk, I will give an overview of research using a unified parameterization approach, the eddy-diffusivity/mass-flux (EDMF) framework, to describe small-scale turbulent motion and cloud properties in a statistically consistent manner. This work involves observational characterization of cloudy boundary layers across spatial and temporal scales to quantify variability as well as a variety of simulation techniques from fine-scale, turbulence-resolving large eddy simulations to coarse-gridded general circulation models.

In this talk, we introduce a correlation robust model for the influence maximization problem. Unlike the classic independent cascade model, this model’s diffusion process is adversarially adapted to the choice of seed set. More precisely, rather than only the independent coupling of known individual edge probabilities, we now evaluate a seed set’s expected influence under all possible correlations—specifically, the one that presents the worst-case. We show that any seed set’s worst-case expected influence can be efficiently computed, and though optimizing the worst-case (over seed sets) is NP-hard, a (1 − 1/e) approximation algorithm can be obtained. We provide structural insights from the model and contrast it with the independent cascade model. We discuss how the proposed model can be extended to optimize other objectives by controlling for conservatism using a mixture of the independent and the worst-case distributions or by incorporating risk criterion in choosing the seed set. Finally, we provide insights from numerical experiments to illustrate the usefulness of the model.

Fueled by recent technological advances, Artificial Intelligence (AI) / Machine Learning (ML) is being introduced to safety- and security-critical applications like defense systems, financial systems, and autonomous machines. AI/ML components can be used either for processing input data and/or for decision making. The response time and success rate demands are very high and this means that the deployed training algorithms often produce complex models that are not readable and verifiable by humans (like multi-layer neural networks). Due to the complexity of these models, achieving complete testing coverage is in most cases not realistically possible. This raises security threats related to the AI/ML components presenting unpredictable behavior due to malicious manipulation (backdoor attacks). This seminar presents ongoing work to develop a systems engineering methodology based on established security principles like zero-trust and defense-in-depth to help prevent and mitigate the consequences of security threats including ones emerging from AI/ML-based components.

Clouds play a critical role in both weather and climate in Earth’s atmosphere. Cumulus clouds that grow into cumulonimbus clouds are the key building blocks for weather phenomena on most spatial scales. For example, the evolution of thunderstorms, tropical cyclones, and the Madden-Julian oscillation all directly impact DoD operations and all are intertwined with the internal dynamics of clouds and their updrafts. This talk will briefly overview the limitations faced by meteorologists in representing cumuliform clouds in weather and climate models as well as progress made toward understanding the internal dynamics of these clouds so that their representation in numerical models of the atmosphere can be further improved.

Machine learning (ML) suffers from a persistent and critical flaw: adversarial examples. Many new forms of adversarial example attacks have been invented and many narrow defenses have been proposed. Unfortunately, no defensive approach can withstand current attacks. We hypothesize that ML model robustness can be improved with approaches that delineate the data-point-sparse latent space between data-dense regions of a model’s classification space as a barrier class. We introduce one such defense, PadNet, that builds a barrier class using a combination of training samples that mix multiple classes together. It leverages this barrier class to separate decision boundaries between benign classes with regions of padding. PadNet then implements a gradient regularization strategy that penalizes gradients in the direction of the barrier class, causing the decision boundary to draw tighter around training samples. We evaluate PadNet against a sampling of the most effective state-of-the-art attacks, demonstrating that it offers more robustness and reliability compared to current defenses.

Acoustic scattering by the seafloor enables remote sensing of seafloor textures, and material properties, but is also a significant source of false alarms for active sonar detection systems used by the Navy. In this talk, an overview is given of what the acoustic scattering problem is and how my own work fits into the larger problem. Specifically, I will detail my work in modeling and measuring acoustic scattering from complex seafloor environments such as rocky outcrops and rough, layered sedimentary seafloors. The application of sonar, as well as the methods used to solve problems, touch on many departments at NPS, and these will be highlighted throughout the work.