This PDF file contains the front matter associated with SPIE Proceedings Volume 11746, including the Title Page, Copyright information, and Table of Contents.

The self-driving car industry has made great autonomy advances, but mostly for well-structured and highly predictable environments. In complex militarily-relevant settings, robotic vehicles have not demonstrated operationally relevant speed and aren’t autonomously reliable. While vehicle platforms that can handle difficult terrain exist, their autonomy algorithms and software often can’t process and respond to changing situations well enough to maintain necessary speeds and keep up with soldiers on a mission. DARPA’s Robotic Autonomy in Complex Environments with Resiliency (RACER) program aims to make sure algorithms aren’t the limiting part of the system and that autonomous combat vehicles can meet or exceed soldier driving abilities. RACER will demonstrate game-changing autonomous UGV mobility, focused on speed and resiliency, using a combination of simulation and advanced platforms. It tests algorithms in the field at DARPA-hosted experiments across the country on a variety of terrain. The program provides UGV platforms that research teams can use to develop autonomous software capabilities through repeated cycles of tests on unstructured off-road landscapes. Goals include not only autonomy algorithms, but also creation of simulation-based approaches and environments that will support rapid advancement of self-driving capabilities for future UGVs.

The Fiscal Year 2019 National Defense Authorization Act (NDAA) established the National Security Commission on Artificial Intelligence (NSCAI) as an independent Commission to “consider the methods and means necessary to advance the development of artificial intelligence, machine learning, and associated technologies by the United States to comprehensively address the national security and defense needs of the United States.”

Introduction to SPIE Defense and Commercial Sensing conference 11746: Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications III

BrainSTORM examines the possibility that the application of AI to STORM, a campaign-level combat model, could reveal novel military Courses of Action (COAs) that catalyze decisive advantages for the U.S. Department of Defense. BrainSTORM is attempting a proof-of-concept demonstration that applies Deep RL to multiple domains of STORM’s campaign in a series of vignette problems of gradually increasing size and complexity. Project milestones examine the potential of tandem human-machine COA development methods similar to the centaur approach of Freestyle Chess. Progress builds toward a live wargame tournament pitting centaur teams - non-expert humans augmented with BrainSTORM - against expert human teams who are not assisted by AI. BrainSTORM’s objectives are to develop limited and credible insights that enhance human understanding and intuition about COAs in war.

This paper provides a perspective on historical background, innovation and applications of Artificial Intelligence (AI) and Machine Learning (ML), data successes and systems challenges, national security interests, and mission opportunities for system problems. AI and ML today are used interchangeably, or together as AI/ML, and are ubiquitous among many industries and applications. The recent explosion, based on a confluence of new ML algorithms, large data sets, and fast and cheap computing, has demonstrated impressive results in classification and regression and used for prediction, and decision-making. Yet, AI/ML today lacks a precise definition, and as a technical discipline, it has grown beyond its origins in computer science. Even though there are impressive feats, primarily of ML, there still is much work needed in order to see the systems benefits of AI, such as perception, reasoning, planning, acting, learning, communicating, and abstraction. Recent national security interests in AI/ML have focused on problems including multidomain operations (MDO), and this has renewed the focus on a systems view of AI/ML. This paper will address the solutions for systems from an AI/ML perspective and that these solutions will draw from methods in AI and ML, as well as computational methods in control, estimation, communication, and information theory, as in the early days of cybernetics. Along with the focus on developing technology, this paper will also address the challenges of integrating these AI/ML systems for warfare.

As the Department of Defense (DoD) looks to exploit and scale Artificial Intelligence (AI) capabilities across the warfighting domains, the Army plans to integrate advanced features into many of its combat systems. The benefits of cloud technologies offer promising solutions to these needs. While cloud-based AI-enabled capabilities leverage flexibility, common interfaces, and virtually infinite scale of resources, they suffer from their lack of proximity to the tactical edge. Tactical AI-enabled systems cannot reliably leverage advantages provided by cloud resources due to limited standardized practices for integration of on-premise/edge systems required by the deployed military. Future high-intensity conflict will be fought in a degraded, denied, intermittent, and lowbandwidth (DDIL) digital environment. As a result, tactical AI-enabled systems will be required to operate in a scenario where high speed, reliable cloud access is unavailable. This paper proposes a hybrid-cloud architecture that leverages resources of the cloud, when available, while also maintaining the capability to retrain tactical AI models in the field environment, using on-site computation and storage. The hybrid cloud construct consists of tactical cloud nodes that reside in closer proximity to AI-enabled systems at the edge. They may retain connectivity to the enterprise cloud yet have the ability to provide the common AI development platform and tool sets to support continuous integration, delivery, and deployment. Thus, its ultimate objective is to enable the seamless and expeditious operation of a distributed AI development environment for the Army and DoD that bridges the tactical edge and enterprise cloud.

This paper presents how the combination of our Distributed InteRactivE C2 Tool (DIRECT) and Multi-Domain Adaptive Request Service (MARS) exploits underutilized resources through distributed adaptation of plans across domains. Deliberate planning processes, especially in the military, tend to be slow and unresponsive. Moreover, the introduction of more flexible assets such as multi-role aircraft introduces latent capacity that is often not exploited due to lack of flexible planning processes, thereby representing a significant opportunity to revolutionize the current system. We seek to overcome these challenges by enabling planners to respond to new requests during execution, through a semi-automated, distributed process that quickly generates options for adapting plans while meeting existing commitments, and presents them for human review. To accomplish this, we infer task state from reported mission states to simplify the manual process of tracking tasks and ensure that the adapted plan incorporates incomplete tasks but does not replan completed tasks. Our dynamic replanner generates options quickly, e.g., 316 seconds to adapt a plan with 345 missions to incorporate 1000 new tasks. This significantly increases utilization of resources, with 60%-70% of imagery requests for battle damage assessment being satisfied by multi-role fighters already flying. Finally, we provide options in context of the existing plan through adaptive option ranking that promotes options that meet operator preferences as judged from abstract evaluation factors designed to apply across different domains. The ranking achieves 80% accuracy for predicting the top option, presenting the preferred option to the operator the vast majority of the time.

Machine Learning (ML) algorithms and Artificial Intelligence (AI) are now regarded as very useful for data-driven applications including resilient multi-domain operations. However, ML algorithms and AI systems can be controlled, dodged, biased, and misled through flawed learning models and input data, they need robust security features and trust. Furthermore, ML algorithms and AI systems add challenges when we have (unlabeled/labeled) sparse/small data or big data for training and evaluation. It is very important to design, evaluate and test ML algorithms and AI systems that produce reliable, robust, trustworthy, explainable, and fair/unbiased outcomes to make them acceptable and reliable in mission critical multi-domain operations. ML algorithms rely on data and work on the principle of ``Garbage In, Garbage Out," which means that if the input data to learning model is corrupted or compromised, the outcomes of the ML/AI would not be optimal, reliable and trustworthy.

Recent breakthroughs and rapid progress in AI will impact, if not transform, every mission. JHU/APL developed an AI Technology Roadmap to guide the Laboratory’s contributions to the critical challenges the nation will face developing and implementing intelligent systems for these missions over the coming decades. We began this exercise by describing a series of envisioned futures for intelligent systems across sea, land, air, space and information, and examined them to identify the common AI technology vectors needed to achieve each vision: (1) Autonomous Perception, describing the path to intelligent systems that perceive in the context of the extreme uncertainty and complexity of the real world; (2) Superhuman Decision-Making and Autonomous Action, to realize the potential for intelligent systems to reason over more information than any team of analysts or operators and act in ways systems under manned control cannot; (3) Human-Machine Teaming at the Speed of Thought, to ensure humans can stay involved at speed and scale; and of particular importance for national security applications, (4) Safe and Assured Operation, so these systems can be trusted to stay true to commander’s intent, in adversarial and sensitive contexts. Each technology vector is aligned with a targeted goal, and with each goal we provide a roadmap in the form of near-, mid-, and long-term AI advances critical to reaching the goal. This paper describes the JHU/APL AI Technology Roadmap and presents key examples of recent progress and forward-looking research and exploratory development along each vector.

Challenges arise when sources of data are needed to provide information to teams of humans in multi-domain battle. Some critical challenges include the dependence and inter-dependence of heterogeneous devices and the uncertainty of information (UoI) obtained from these devices. UoI significantly affects the decision-making process and humans rely on underlying reasons for uncertainty in making decisions that rely on devices and data from these devices. The LRM Method is an excellent tool utilized to assist a decision-maker in support of military relevant operations. Previously, the LRM algorithm optimized the UoI and determined the decomposition of the UoI into its various elements via MATLAB. In recent work, the LRM algorithm is performed via Java. Although both algorithms compute the UoI to an error of within a negligible tolerance, there are small disparities in the decomposition of the UoI. Decomposing the UoI into its various parts is the main benefit of utilizing the LRM Method. These differences can have a significant impact on the decision-maker and the decision-making process.

Graphical model, diffusion equation network dynamics, and manifold learning are different subareas in machine learning and network science. In this paper, combining the ideas from these subareas, we propose and implement graphical model and diffusion equation based manifold learning techniques for sensor data processing. We show that the graphical model and diffusion equation combined manifold learning can be used to perform data processing for one, two, and three-dimensional sensors. Experiments show that this manifold learning approach can solve many sensor data processing problems including radio frequency signal processing, image processing (computer vision), and three-dimensional lighting detection and ranging (LIDAR) processing problems better than some traditional methods.

Data reduction techniques are becoming increasingly important in modern data analysis problems{especially in fields involving big data. Although not quite on the same scale as high velocity and volume big data problems, remote sensing analysis on hyperspectral imagery (HSI) can suffer from the increase in dimensionality provided by this sensing modality. Common hyperspectral sensors divide the sensed optical bandwidth into as many as 270 different channels. Depending on the specific bandwidth and application, some channels may be unusable due to water absorption, while others may provide little-to-no discriminative information when considering their impact on algorithm performance. With this in mind, there are several techniques devised to subsample the original number of bands to a more manageable size allowing for increased algorithm performance both in classification accuracy and computation time. Herein, we propose a neural network-based band selection technique that seeks to generalize existing autoencoder-based methods and experiment with tracking input contributions throughout the network. We explore the ability to perform task-specific band selection by applying this method to a general multilayer perceptron architecture. We apply our proposed band selection methodology to a novel HSI dataset captured by the U.S. Army ERDC and compare performance to alternative band selection algorithms.

Using a semantic hierarchy as a labeling scheme can provide object detection systems with a more robust expert knowledge of the relationships between object classes. This knowledge can be used to improve object class prediction in cases where an object detector encounters an object of a class upon which it was not trained, known as zero-shot object detection or open-set recognition. Datasets which are useful for a particular application, domain, or task may not have their object labels organized into an appropriate semantic hierarchy. For example, the Scene UNderstanding (SUN¹) Database has image scenes organized into a hierarchy, but no such organization exists with the object labels. Objects in the images of this dataset were annotated in a crowd-sourced manner which allowed annotators to define the polygons which bound the objects as well as assign the labels. The challenge taken up by the method presented in this paper was to take the original object labels of the SUN Database and create a semantic hierarchy such that each child-parent pair of object classes demonstrates an “IS-A” relationship. By associating common labels within the dataset with the most relevant and fine-grained WordNet² synonym, this approach resulted in a multi-layered semantic hierarchy for SUN Database object labels. The product is a tree-structured graph where each node is a WordNet synonym of the original label and a node’s parent is determined by its WordNet hypernym. Other ontological frameworks, such as Basic Formal Ontology³ and the Operational Environment Ontology Suite,⁴ are also discussed.

The US Army Combat Capabilities Development Command (DEVCOM) is sponsoring a unique international prize competition focused on identifying innovative, relevant technologies in Artificial Intelligence (AI) from non-traditional/small Business technology developers in Europe and the Middle East. The event – xTech Global AI Challenge – is an extension of the US Army’s very successful Expeditionary Technology Search (xTechSearch) program, expanding the concept for the first time to a Global audience. The purpose behind this AI Challenge is to help the DoD (and coalition partners) rapidly and effectively harness massive amounts of data in resource-constrained environments on current and future battlefields. The RFP will open to international industry in March 2021, and the pitch event will take place the on 13-15 July 2021 in London, UK in July 2021 , at the Innovation Hub (iHUB) DoD space at Imperial College London (ICL) hosted by U.S. Army and UK Defense & Security Accelerator (DASA), with support from the DEVCOM Army Research Laboratory and the Army Artificial Intelligence Task Force. This talk will highlight the xTech Global event, as well as the DEVCOM Atlantic organization and their role in extending the US Army Science and Technology mission abroad. Prize competitions such as this are just one example of the many ways the US Army works with technology developers around the world to advance scientific discovery and transition of ground-breaking technologies into society.

This paper describes an unsupervised machine learning methodology capable of target tracking and background suppression via a novel dual-model approach. “Jekyll” produces a video bit-mask describing an estimate of the locations of moving objects, and “Hyde” outputs a pseudo-background frame to subtract from the original input image sequence. These models were trained with a custom-modified version of Cross Entropy Loss. Simulated data were used to compare the performance of Jekyll and Hyde against a more traditional supervised Machine Learning approach. The results from these comparisons show that the unsupervised methods developed are competitive in output quality with supervised techniques, without the associated cost of acquiring labeled training data.

The next generation of multi-domain airborne platforms will provide military operators with unparalleled sensor data streams spanning video, radar, and other sensor inputs. These expanded sensor capabilities will substantially increase access to critical, near-real time surveillance. However, the process of interpreting video feeds places a significant burden on intelligence operators, a demand that can be addressed by AI-based algorithms. AI-based tools complement the aerial footage processing tasks performed by full motion video analysts. In this work, we introduce a new aerial pattern of life dataset and describe our latest algorithmic developments, which uses deep learning to gain an understanding of a scene’s patterns of life. This approach allows anomalies, outliers from standard patterns of life, to be identified using supervised and unsupervised learning approaches. Herein, we describe our deep learning models and our corresponding microservices software architecture. The patterns of life and anomaly detection performance is measured through analysis of video from this new remotely piloted aerial system (RPAS) flight campaign dataset.

In many real-world multi-domain applications, if there is an opportunity to sense the opponent's strategy from previous rounds, an agent can exploit its opponent in payoffs by playing a specific Best Response (BR) strategy. The optimal way to exploit the opponents in such contests is to learn specific BR strategies for each candidate's strategies. However, in a large strategy space, learning BR policies for all the strategies is infeasible. We need to find a reasonable point to learn some BR strategies to exploit the opponent strategies, which motivates us to propose a novel method for finding better agent strategies while facing a specific set of opponents in repeated games. Our method is called Clustered Double Oracle Empirical Game-Theoretic Analysis (CDO-EGTA), which builds upon the classic Double Oracle (DO) framework based on Deep Q-Network (DQN) and clustering methods. Empirical results show that our method outperforms the current state-of-the-art methods in terms of regret.

This paper proposes a novel region-based structure measure for object proposal ranking. It is able to efficiently reduce redundant object proposals and highlight dominant objects in an image. The computation of this measure is fully unsupervised, without any image level annotation or any visual semantics labeling. In this work, a new set of heuristic rules are introduced to indicate regions that may contain objects. The distinctiveness of a proposal region is assessed based on its structural uniqueness, structural distributions and deformable shapes. A scoring function is then constructed to combine these multi-view rule-based assessments into a single object score. Furthermore, a rank-recall optimization is proposed to optimize the scoring function for proposal ranking. The final optimized ranking significantly reduces the number of object proposals while maintaining potential object regions. Promising results show that the proposed ranking method simultaneously reduces proposals and highlights dominant objects.

The complexity of modern military operations create a demand for efficient collaborative decision making and problem solving. Additionally, as military units operate in increasingly dynamic environments, the ability to respond to changing circumstances becomes paramount for mission success. An effective response rests on correct dissemination and transfer of information across the command and control structure, and thus is critically linked to the network of human interactions. In this paper, we take an agent-based modeling approach to collective problem solving. We investigate three key factors affecting the performance in collaborative environments: (1) the structure of network used to share information between agents, (2) the search strategies adopted by agents, and (3) the complexity of problems facing the group. In particular we study how the trade-off between exploitation of known solutions and exploration for novel ones is related to the efficiency of collective search. Additionally we consider the role of agent behavior: propensity for risk-taking and trustworthiness, as well as the dynamic nature of social connections. Finally, we outline the directions for future work regarding the efficiency of problem solving on military-like command and control structures.

Warfighters at the tactical edge do not have “Big Compute” capability and are bandwidth constrained. This significantly affects capability, workload and utility of Intelligence, Surveillance and Reconnaissance (ISR) at the edge. Artificial Intelligence (AI) has the potential to improve quality of information and reduce workload if we optimize AI/ML (Machine Learning) techniques for the constrained tactical environment. For several years, we focused on sensor and data interoperability. Leveraging contested urban environment experiments, and our work with Canada, we cracked the “data interoperability nut” across the systems provided by Canada, the United Kingdom, Australia, and New Zealand. The increased data interoperability made it clear that the link between sensor interoperability and data reasoning is a missing component of ISR at the edge. Data overload for the tactical operators was almost immediate and overwhelming. The success in interoperability highlighted failures in data fusion, filtering, and presentation. Increased data flow does not equal increased situational awareness. CATE is a collaborative effort to develop a prototype AI/ML architecture and framework that enables simple, rapid integration of collaborative, multi-agent AI technology into the Processing, Exploitation and Dissemination (PED) chain at the edge, sensor, and tactical PED level. CATE imagines a soldier on patrol in a “smart city”. In the background, collaborative AI agents scan city cameras, review patterns of life, providing an AI enabled over watch. The AI agents determine there is a threat and alert the soldier, who never had to look down at a screen and take his eyes off his immediate surroundings.

Continued advancements in adversarial attacks have crippled neural network performance. These small pixel perturbations can go undetected and cause networks to misclassify with high confidence. The motivation for this paper was to investigate how various sensor modalities and network models respond to adversarial attacks. It is important to realize that the large diversity in neural network architectures makes it difficult for any analytical conclusions to be made that generalize across any given neural network. For this reason, we share the statistical analyses performed which could be applied to any network under review. General observations gained from this analysis are also shared which indicated that network classification accuracy is not just a function of the network model but the data as well.

In adversarial multi-agent paradigms, it is often difficult to describe what is learned by reinforcement-learning agents and to measure the robustness of the learned policies. To achieve a better view, we explore the effect of manipulating agent capabilities and policies in a predator-prey pursuit task. In these experiments, we trained a single prey using multi-agent reinforcement learning with three slower predators, then tested the prey against three faster predators using fixed “interceptor” strategies (head to the closest possible intersection with the prey assuming the prey maintains its current velocity) instead of their learned policies. While the prey’s performance was impressive under these novel conditions, it varied widely. Initial locations and velocities (randomized during training and testing) were limited in explaining differences in prey’s performance across test conditions. Nevertheless, visual inspection indicates that more successful prey quickly begin a circling pattern, whereas less successful prey often become cornered and double back into predator collisions. To quantify this behavior, we computed windowed entropy measures of the prey’s angle relative to the arena origin to show when an agent transitioned in and out of this unsuccessful behavior. Ultimately, these transitions suggest that circling is triggered by coming close to the center of the arena. By varying agents’ capabilities and predator policies upon evaluation, we achieve a more comprehensive view of the prey’s learned policy, and we suggest that these windowed entropy measures, along with correlations between entropy and performance, result in a quantification of the learned policy.

Increasingly cyber-attacks are sophisticated and occur rapidly, necessitating the use of machine learning techniques for detection at machine speed. However, the use of machine learning techniques in cyber security requires the extraction of features from the raw network traffic. Thus, subject matter expertise is essential to analyze the network traffic and extract optimum features to detect a cyber-attack. Consequently, we propose a novel machine learning algorithm for malicious network traffic detection using only the bytes of the raw network traffic. The feature vector in our machine learning method is a structure containing the headers and a variable number of payload bytes. We propose a 1D-Convolutional Neural Network (1D-CNN) and Feed Forward Network for detection of malicious packets using raw network bytes.

As autonomous driving technologies evolve, an in-vehicle network connecting numerous sensors, actuators, and electric control units (ECUs) has become increasingly important and has led to the critical need for ensuring the security of these networks. These ECUs and vehicle components in an in-vehicle network require a more reliable and fast data transport protocol than those in ordinary computer communication. To meet these requirements, the controller area network (CAN) protocol is used in which a CAN frame containing a small payload related to the state and control of a vehicle is sent. Because the CAN protocol broadcasts unencrypted messages to the bus, it is exposed to many security threats and vulnerabilities. In particular, a network can be easily compromised by attacks such as denial-of-service (DoS), fuzzy attacks, and spoofing as long as the attacker can access the CAN network. In this study, we develop a novel deep convolutional neural network (DCNN)-based attack detection technique for CAN. Specifically, we use two key characteristics that can be obtained by observing CAN traffic flows. The first is the statistical distribution of CAN frame appearances per unit time, and the second is the average interarrival time (IAT) of the CAN frames. These characteristics are measured at different levels of time granularity and are aggregated to constitute traffic samples for DCNN-based attack detection. By processing these samples and inputting them into the DCNN, we can determine the presence or absence of an attack during each time interval in real time. Because the proposed method utilizes statistical characteristics at different levels of time granularity, it can effectively detect attacks performed in both wide and narrow time intervals.

Deep Learning (DL) is becoming a popular paradigm in a broad category of decision systems that are crucial to the well-being of our society. Self-driving vehicles, online dating, social network content recommendation, chest X-Ray screening, etc. are all examples that show how the quality of our lives is tied to the decisions of these systems. We must take into account that these systems may be gamed to make favorable decisions for unqualified instances by malicious actors. For instance, if a self-driving car's traffic-sign detection model can classify a traffic stop sign as speed-limit if the pattern that triggers the faulty behavior is present. Our initial investigation result show, given we can generate/access a rich and high-quality dataset of random images, we may be able to build meta-models that can distinguish the poisoned/clean models with acceptable performance.

Generative Adversarial Network (GAN) is regarded as a class of generative models that relies on unsupervised machine learning. In this paper work, we design and evaluate a novel optimized GAN model with anomaly detection algorithm and compare it with a base signature anomaly GAN detection model. Typical anomaly detection algorithms that depend on the generative hostile network that are expected to produce robust and accurate detection of anomalies in a scalable number of data points, result in decrease in accuracy because of lack of enough data or unfiltered resampling data techniques. Thus, unlike in traditional neural networks for feature identification, our proposed approach leverages anomaly and mining prediction GAN ingestion with unique monitoring engine features for enhancing the identification, system resiliency, and overall performance with weight updates for the discriminator and generator models. Performance evaluation results show that the proposed approach results in better detection over the state-of-the-art predictive botnet detection.

In this study, we approach the problem of image captioning with cycle consistent generative adversarial networks (CycleGANs). Due to CycleGANs’ ability to learn functions to map between multiple domains and use duality to strengthen each individual mapping with the usage of a cycle consistency loss, these models show great promise in their ability to learn both image captioning and image synthesis and to create a better image captioning framework. Historically, cycle consistency loss was based on the premise that the input should undergo little to no change when mapped to another domain and then back to its original; however, image captioning presents a unique challenge to this concept due to the many-to-many nature of the mapping from images to captions and vice-versa. TextCycleGAN overcomes this obstacle through utilization of cycle consistency in the feature space and is, thereby, able to perform well on both image captioning and synthesis. We will demonstrate its capability as an image captioning framework and discuss how its model architecture makes this possible.

Imitation learning has been shown to be a successful learning technique in scenarios where autonomous agents have to adapt their operation across diverse environments or domains. The main principle underlying imitation learning is to determine a state-to-action mapping, called a policy, from trajectories demonstrated by an expert. We consider the problem of imitation learning under adversarial settings where the expert could be malicious and intermittently give incorrect demonstrations to misguide the learning agent. We propose a technique using temporally extended policies called options to make a learning agent robust against adversarial expert demonstrations. Experimental evaluation of our proposed technique for a game playing AI shows that a learning agent using our options based technique can successfully resist deterioration in its task performance as compared to using conventional reinforcement learning, when an expert adversarially modifies the demonstrations either randomly or strategically.

Given the rapid changes in telecommunication systems and their higher dependence on artificial intelligence, it is increasingly important to have models that can perform well under different, possibly adverse, conditions. Deep Neural Networks (DNNs) using convolutional layers are state-of-the-art in many tasks in communications. However, in other domains, like image classification, DNNs have been shown to be vulnerable to adversarial perturbations, which consist of imperceptible crafted noise that when added to the data fools the model into misclassification. This puts into question the security of DNNs in communication tasks, and in particular in modulation recognition. We propose a novel framework to test the robustness of current state-of-the-art models where the adversarial perturbation strength is dependent on the signal strength and measured with the “signal to perturbation ratio” (SPR). We show that current state-of-the-art models are susceptible to these perturbations. In contrast to current research on the topic of image classification, modulation recognition allows us to have easily accessible insights on the usefulness of the features learned by DNNs by looking at the constellation space. When analyzing these vulnerable models we found that adversarial perturbations do not shift the symbols towards the nearest classes in constellation space. This shows that DNNs do not base their decisions on signal statistics that are important for the Bayes-optimal modulation recognition model, but spurious correlations in the training data. Our feature analysis and proposed framework can help in the task of finding better models for communication systems.

In this paper, we propose an energy efficient cooperative monitoring algorithm for detecting and localizing adversarial communications in the presence of malicious reporting, Our solution is based on a combination of a game theory formulation for distributivity managing the monitoring-energy tradeoffs and a novel recursive trilateration algorithm which provides robustness to measurement errors and malicious reporting. Our results show that the energy consumption for monitoring is significantly reduced for our proposed solution compared to the all monitoring benchmark, and that the algorithm shows high robustness to malicious reporting. Our proposed method was able to still correctly localize the adversarial source within a specified tolerance error, and within a prescribed number of monitoring epochs, 83% of the time, when 25% of the monitoring nodes were compromised, reporting maliciously.

Uncertainty represents the quantification of the spread of the distribution of possible ground truths that can be inferred from observed evidence. As such, uncertainty is one of the major factors in determining confidence when making decisions (i.e., uncertainty and confidence are in an inverse relationship). Bayesian statistics and subjective logic provide tools for Artificial Intelligence (AI) to derive uncertainty quantification. These processes require base rates, which are large-population determinations of probabilities that are not contextualized for the specific situation. The AI computes probabilities based upon the specific situation and context in light of historical (or training) data. As more evidence/training data is available for the context, the base rate gets washed out in the probability calculation. For most Army applications, an AI does not act or decide on its own with the rare exception of complete automaticity, but rather in collaboration with at least one human user. In this paper, we propose that the ways AI represents uncertainty ought to be optimally aligned with human preferences to provide best possible human-AI collaborative performance. Exploring this topic requires human-subjects experimentation to test how well users understand different representations of uncertainty that include base-rate information, which quantifies belief in predictions. Variations of these experiments could include different types of training to interpret uncertainty representations.

This presentation will describe the Combat Capabilities Development Command (DEVCOM) Armament Center’s research on human behavior for machine learning and development of artificial intelligence aids. Human data collection for the development of artificial intelligence aid development can be roughly categorized into three categories. The first can be described as collection of human behavioral data in order to base artificial intelligence on natural intelligence. The second is the work in turning artificial intelligence into artificial intelligence aids, with a focus on the examination of the interaction between human and artificial intelligence aid. The third is research examining validation—confirming that the artificial intelligence aid actual aids the natural intelligence. For the past few years the DEVCOM Armament Center’s Tactical Behavior Research Laboratory has engaged in research efforts aimed at generating human data sets for machine learning analyses while conducting testing, evaluation and validation of notional artificial aids. Several efforts across the three categories will be described. These include human data gathering efforts for a targeting prioritization, human threat identification in an urban environment, and electrophysiological interfaces. The data from these efforts are inputted into appropriate machine learning methodologies from which algorithms underlying artificial intelligence aids may be derived. Specialized facilities (the Experimental Verification and Validation Assessment Lab) for gathering these data will be described. Finally, the presentation will include an overview of the lab’s processes and pipelines from its initial data gathering and algorithm development to its testing and evaluation and ultimately, to its verification and validation of artificial intelligence aids.

Future Multi Domain Operation (MDO) wargaming will rely on Artificial Intelligence/Machine Learning (AI/ML) algorithms to aid and accelerate complex Command and Control decision-making. This requires an interdisciplinary effort to develop new algorithms that can operate in dynamic environments with changing rules, uncertainty, individual biases, changing cognitive states, as well as the capability to rapidly mitigate unexpected hostile capabilities and exploit friendly technological capabilities. Building on recent advancements in AI/ML algorithms, we believe that new algorithms for learning, reasoning under uncertainty, game theory with three or more players, and interpretable AI can be developed to aid in complex MDO decision-making. To achieve these goals, we developed a new flexible MDO warfighter machine interface game, Battlespace, to investigate and understand how human decision-making principles can be leveraged by and synergized with AI. We conducted several experiments with human vs. random players operating in a fixed environment with fixed rules, where the overall goal of the human players was to collaborate to either capture the opponents’ flags or eliminate all of their units. Then, we analyzed the evolution of the games and identified key features that characterized the human players’ strategies and their overall goal. We then followed a Bayesian approach to model the human strategies and developed heuristic strategies for a simple AI agent. Preliminary analysis revealed that following the human agents’ strategy in the capture the flag games produced the greatest winning percentage and may be useful for gauging the value of intermediate game states for developing the coordinated action planning of reinforcement learning algorithms.

Human-in-the-loop (HITL) is the process of combining the power of a machine or computer system and human intelligence to develop human-aware machine learning (ML) models. The HITL process creates a continuous feedback loop between human and machine, enabling the trained model to continuously improve as edge cases present themselves without the need to fine-tune the model from scratch. Several advantages of utilizing HITL ML systems are to avoid bias, ensure consistency and accuracy, improve efficiency, and provide transparency. However, adding human involvement also invites human mistakes. Occasionally, a HITL system may actually degrade the algorithm rather than improve it: mislabeling an object in an object detection algorithm, incorrectly scoring an algorithm’s output, making misclicks and typos, and other human errors cause the HITL to make a mistake based on the facts presented to it. These errors being made by the user, intentional or not, can be considered a form of data poisoning. To understand the effects of a HITL’s choices on an ML model, several pieces of information during the HITL process can be observed, i.e.,: the time taken by the user to provide input on an output or on a specific object class, as well as an evaluation of the consistency of submitted valid inputs, among other factors. Information extracted from the HITL’s decision-making process can provide insights into whether poor choices are being made by the user (i.e., data poisoning) and identify where, when, and why these choices are being made. Many state-of-the-art models can be utilized for this work, such as ResNet-50, DarkNet-53, Xception, among others. However, for this work, we are less focused on the model being used and more focused on the procedure for tracking HITL performance to maximize model improvement. Nevertheless, this work will consider a pretrained model, though the approach will be model agnostic. The dataset used in this research is the publicly available “Flowers Recognition” dataset available on Kaggle.¹

Localization, tracking, and classification requires precise recovery of the target signal amidst a sea of environmental noise and reflections from terrain and buildings. In this work, we use machine learning (ML), specifically generative-adversarial networks (GANs), to remove a range of synthetic additive noise and convolutional reverberations from a well-defined subset of sounds, namely English speech. Unlike simple denoising autoencoders, GANs attempt to generate realistic solutions, suppressing the production of unresolved (i.e., blurry) mixed inferences. We demonstrate that our models can dynamically resolve varying amounts of noise and convolution, which will be important in the field, where the amount and type of signal degradation will generally be unknown. While we focus only on speech here, ML-based deconvolution can also be applied to the restoration of images, radar, radio, and acoustic and seismic sensing distorted by weather, interference, and reflections.

Multi-Agent Reinforcement Learning (RL) algorithms have recently demonstrated their ability to discover/develop new strategies to solving problems. This work aims to apply this same methodology to solving convex optimization problems in an attempt to train agents to outperform popular optimizers. This work will examine methods to frame the problem to the RL agents which encourage rapid convergence. Additionally we will discuss the various ways to represent the environment or state that the agents will operate in, as well as the action space and rewards.

This work presents a novel method to reconstruct ultra-wideband radar signals over their entire bands when only a portion of the spectrum is available. The inherent sparse configuration of sensing radars is exploited to develop a robust wideband synthesizing framework. The proposed approach consists of the following two steps: 1) A radar signal model is developed which allows for a sparse representation of the target’s signature over wide bandwidths. 2) A flexible statistical model is introduced to solve the sparse model of Step 1 in a Bayesian framework. By imposing Gamma-Gaussian sparsity promoting priors on the target’s signature, the model improves the sensing accuracy. In addition, by introducing a Gaussian mixture model of noise, any non-deterministic density of noise or jamming can be accurately approximated. This significantly increases robustness in complex environments. The proposed approach is verified through the development of an ultra-wideband (12–110 GHz) radar system consisting of canonical-spherical targets. Monte Carlo simulations are carried out to compare the performance of the proposed statistical method when the data missing rate is 25% with other methods in the literature. It is shown that the proposed model has closed-form solutions and is robust to complex environments with satisfying performance. As the noise distribution violates the single Gaussian assumption, other relevant methods in the literature fail to recover the model with a large RMS error (< 0.50). In contrast, our method has a maximum RMS of only 0.11.

The International Civil Aviation Organization’s 2020 mandate for all aircraft to have Automatic Dependent Surveillance-Broadcast (ADS-B) equipment installed poses a unique challenge to the military. With the use of cheap sensors, adversaries now can collect real-time data on flights transporting troops. In this paper, we outline a statistical model that can be leveraged both defensively and offensively. Our model, a time-series spatial process model with a first-order Markov assumption, can be used to develop a flight path that could help a military aircraft seem like ordinary air traffic to an observer. This would prevent adversaries from identifying and tracking these flights as easily. The model can also be used to predict the destination of flights based on their observed location. These real-time flight predictions use Bayesian updating to improve their accuracy as more information about the flight is observed.

Vehicle maneuverability is often supported in low-light scenarios through infrared (IR) imagery. However, if the imagery contains little temperature gradient, the raw images are less applicable. In order to maximize image effectiveness, a genetic algorithm (GA) is employed to explore various contrast enhancement operators to determine an optimal sequence of contrast enhancements. We propose a new image quality evaluator that incorporates the performance of a deep learning-based object detector and considers image spatial context through cell-structured configurations. The proposed technique is assessed both qualitatively and quantitatively for the task of maneuverability hazard detection.

To function at the same operational tempo as human teammates on the battlefield in a robust and resilient manner, autonomous systems must assess and manage risk as it pertains to vehicle navigation. Risk comes in multiple forms, associated with both specific and uncertain terrains, environmental conditions, and nearby actors. In this work, we present a risk-aware path planning method to handle the first form, incorporating perception uncertainty over terrain types to trade-off between exploration and exploitation behaviors. The uncertainty from machine learned terrain segmentation models is used to generate a layered terrain map that associates every grid cell with its label uncertainty among the semantic classes. The risk term increases when differently traversable semantic classes (e.g., tree and grass) are associated with the same cell. We show that adjusting risk tolerances allows the planner to recognize and generate paths through materials like tall grass that historically have been ruled out when only considering geometry. Utilizing a risk-aware planner allows triggering an exploratory behavior to gather more information to minimize uncertainty over terrain categorizations. Most existing methods for incorporating risk will avoid regions of uncertainty, whereas here the vehicle can determine if the risk is too high after new observation/investigation. This also allows the autonomous system to decide to ask a human teammate for help to reduce uncertainty and make progress towards goal. We demonstrate the approach on a ground robot in simulation and in real world for autonomously navigating through a wooded environment.

Cybersecurity of autonomous vehicles is a pertinent concern both for defense and also civilian systems. From self-driving cars to autonomous Navy vessels, malfunctions can have devastating consequences, including losses of life and infrastructure. Autonomous ground vehicles use a variety of sensors to image their environment: passive sensors such as RGB(-D) and thermal cameras, and active sensors such as LIDAR, radar, and sonar. These sensors are either used alone or fused to accomplish the basic mobile autonomy tasks: obstacle avoidance, localization, mapping, and, subsequently, path-planning. In this paper, we will provide a qualitative and quantitative analysis of the effect of perturbed sensing capability of depth sensing, focusing on LIDAR, and the subsequent effects on navigation and path planning in the presence of obstacles. Aspects that will be investigated include complexity of the perturbation and effect on the autonomous operations. This work will lay a foundation for developing robust autonomy algorithms that are secure against possible degraded or inoperable sensors.

Multi-agent reinforcement learning (MARL) has transformed research and development in robotics especially for navigation purposes. Using deep learning, multi-agent coordination, reinforcement learning can solve critical tasks such as finding shortest paths or avoiding obstacles with exceptional speed. However, given a LiDAR point cloud structure, performing such tasks directly using MARL, can be computationally expensive and cumbersome due to massive nature of 3D point clouds. In this work, we leverage 2D (MARL), path planning and obstacle avoidance to obtain a low latency navigation in 3D though a correspondence 2D floor plan. We change the environment dynamically with adversarial threats such as fire or leakage -- the RL agents can quickly find a new path given the learned environment. The method shows strong performances in both throughput and latency.

Our goal is to develop stable, accurate, and robust semantic scene understanding methods for wide-area scene perception and understanding, especially in challenging outdoor environments. To achieve this, we are exploring and evaluating a range of related technology and solutions, including AI-driven multimodal scene perception, fusion, processing, and understanding. This work reports our efforts on the evaluation of a state-of-the-art approach for semantic segmentation with multiple RGB and depth sensing data. We employ four large datasets composed of diverse urban and terrain scenes and design various experimental methods and metrics. In addition, we also develop new strategies of multi-datasets learning to improve the detection and recognition of unseen objects. Extensive experiments, implementations, and results are reported in the paper.

A decision is a conclusion reached after considering information, while decision making is the process of making choices after assessing alternatives with the information gathered. Most decisions made lead to an action or a choice. However, this process is often interactive and iterative in nature. Interaction can be from people, agents, or both. The interaction can take the form of additional data, different criteria, and changes to the goals for the decision. A single decision may not be the final decision but made at a single point that flows into another. Thus, iterations can initiate this flow of decisions at various points and in a cycle. In the decision cycle the decision may be repeated, as fine tuning of the system or tasks occur. Then there is the ongoing challenge that decisions nor the decision making process is simple nor is it perfect. Many techniques strive to capture the uncertainty that this imperfection creates. In this work we are referencing continued research being conducted under Artificial Reasoning for Uncertainty of Information (UoI). UoI is a reason based approach that builds on the concept of imperfect information. The objective of the UoI research is to represent and present the reasons, causes for specified uncertainty for a task, specifically the decisions for the task. As the UoI research expands to ideally allow greater adaptability, the area of graphs is being explored. Graphs are a decision making tool where connections between inputs (information, criteria, goals, etc…) and outputs (alternatives, choices, …) can be shown and analyzed. Utilizing graphs is also helpful as new capabilities are integrated into the UoI concept. This paper will explore how graphs are and can be used to incorporate new features and new capabilities particularly for selected warfighter functions.

We present an explanation system for applications that leverage Answer Set Programming (ASP). Given a program P, an answer set A of P, and an atom a in the program P, our system generates all explanation graphs of a which help explain why a is true (or false) given the program P and the answer set A. We illustrate the functionality of the system using some examples from the literature.

Explanations are generated to accompany a model decision indicating features of the input data that were the most relevant towards the model decision. Explanations are important not only for understanding the decisions of deep neural network, which in spite of their their huge success in multiple domains operate largely as abstract black boxes, but also for other model classes such as gradient boosted decision trees. In this work, we propose methods, using both Bayesian and Non-Bayesian approaches to augment explanations with uncertainty scores. We believe that uncertainty augmented saliency maps can help in better calibration of the trust between human analyst and the machine learning models.

Knowledge graphs (KGs) provide a useful representation format for capturing complex knowledge about an information domain, with rich logical descriptions available for defining the relationships between entities. Separately, semantic vector spaces (SVSs) capture the relative meanings of terms based on their actual usage within a dataset and allow useful operations for exploring the relationships between these terms. Combining KGs and SVSs via knowledge graph embedding (KGE) enables further analysis tasks to leverage learned semantic vectors to gain additional insights. Therefore, KGE represents an interesting and potentially powerful tool for identifying emergent or unexpected behavior, or for seeking previously unaccounted for relationships, event, and groups. In this work, we report on the state-of-the-art in KGE. We describe the operational benefits that can be gained from this approach and the considerations that apply for observational ontologies that describe a complex, untrusted, time-sensitive, and rapidly-evolving environment. We suggest several promising avenues for future research in this context.

The rapid development in data collecting devices and computation platforms produces an emerging number of agents, each equipped with a unique data modality over a particular population of subjects. While an agent’s predictive performance may be enhanced by transmitting others’ data to it, this is often unrealistic due to intractable transmission costs and security concerns. In this paper, we propose a method named ASCII for an agent to improve its classification performance through assistance from other agents, without sharing proprietary data and model information. The method is naturally suitable for privacy-aware, transmission-economical, and decentralized learning scenarios. The method is also general as it allows the agents to use arbitrary classifiers such as logistic regression, ensemble tree, and neural network, and they may be heterogeneous among agents. We demonstrate the proposed method with extensive experimental studies.

Future Multi-Domain Operations (MDO) will require the coordination of hundreds, even thousands, of devices and component services. This will demand the capability to rapidly discover the distributed devices/services and combine them into different work ow configurations, thereby creating the applications necessary to support changing mission needs. To meet these objectives, we envision a distributed Cognitive Computing System (CCS) that consists of humans and software that work together as a ‘Distributed Federated Brain'. Motivated by neuromorphic processing models, we present an approach that uses hyper-dimensional symbolic semantic vector representations of the services/devices and workflows. We show how these can be used to perform decentralized service/device discovery and work ow composition in the context of a dynamic communications re-planning scenario. In this paper, we describe how emerging analogue AI ‘In Memory' and ‘Near Memory' computing devices can be used to efficiently perform some of the required hyper-dimensional vector computation (HDC). We present an evaluation of the performance of an energy-efficient phase change memory device (PCM) that can perform the required vector operations and discuss how such devices could be used in energy-critical ‘edge of network' tactical MDO operations.

Software Defined Coalition (SDC) is a promising architecture for flexibly controlling network resources in tactical environments. In highly dynamic mobile ad hoc network scenarios, however, the SDC controllers may be fragmented from the mobile nodes they manage rendering impossible their in-time resource reconfiguration and causing severe network performance degradation. To address this challenge, we explore the idea of complementing the SDC controllers with distributed control mechanisms running locally at the mobile nodes in a hybrid architecture design. We then apply learning methods to predict the network dynamics and fragmentation events, so that we can tune the routing policies in time to prevent network failures from happening. We show the benefits of our hybrid approach in a demo emulating a real tactical ad hoc network scenario using the Anglova tactical ad hoc network dataset.

By extending the Software Defined Networking (SDN), the Distributed Analytics and Information Sciences International Technology Alliance (DAIS ITA) https://dais-ita.org/pub has introduced a new architecture called Software Defined Coalitions (SDC) to share communication, computation, storage, database, sensor and other resources among coalition forces. Reinforcement learning (RL) has been shown to be effective for managing SDC. Due to link failure or operational requirements, SDC may become fragmented and reconnected again over time. This paper shows how data and knowledge acquired in the disconnected SDC domains during fragmentation can be used via transfer learning (TL) to significantly enhance the RL after fragmentation ends. Thus, the combined RL-TL technique enables efficient management and control of SDC despite fragmentation. The technique also enhances the robustness of the SDC architecture for supporting distributed analytics services.

Recent years have seen significant advances in artificial intelligence (AI) and machine learning (ML) technologies applicable to coalition situational understanding (CSU). However, state-of-the-art ML techniques based on deep neural networks require large volumes of training data; unfortunately, representative training examples of situations of interest in CSU are usually sparse. Moreover, to be useful, ML-based analytic services must be capable of explaining their outputs. We describe an integrated CSU architecture that combines neural networks with symbolic learning and reasoning to address the problem of sparse training data. We also demonstrate how explainability can be achieved for deep neural networks operating on multimodal sensor feeds. The work focuses on real-time decision making settings at the tactical edge, with both the symbolic and neural network parts of the system --- including the explainabilty approaches --- able to deal with temporal features.

Social network analysis is well-known as a fundamental tool for structural assessment of relationships at scale. The overwhelming convention in this field is to model relations based on positive factors, often related to commonality between participants, such as mutual friendship, or other characteristics that may positively bind participants. However, this leaves undefined the meaning of absent links between agents, leaving open scope for alternative interpretations. For example, an absent link may be due to individuals being unfamiliar and irrelevant to each other, but it could also be due to negative relations. This context is particularly relevant to adversarial scenarios, where actors, or the groups that they represent, are the focus of analysis. Accordingly, in this paper we consider the importance of explicitly modelling so-called negative ties as a means to provide insight into conflict and adversarial interaction. This subtle redefinition of social network analysis provides an opportunity to gain new perspectives on the threats that can subvert traditional social network analysis due to negativity not being explicitly represented. In this paper we focus on the fundamental issue of how to define negative ties. Generally, these invoke or are the consequence of cognitive dissonance between conflicting parties. We distinguish further between these three domains of negative, network connections and provide illustrations of each type in the context of social conflict. We discuss implications of these various genres of negativity regarding their potential to be used as tools by adversaries and their ability to provide a deeper understanding of human conflict.

We provide a description of the DAIS-ITA experimentation environment with accompanying scenario. These were created to aid our research by providing a common platform upon which different projects can build. The purpose is to speed up the innovation pipeline by more closely linking research with impact. The experimentation environment is based on a 3d virtual world. We provide a variety of options for displaying, moving and interacting with objects within the world. Our focus for these interactions is interactivity, either manual or programmatic, such that DAIS research, experiments, simulations and demonstrations can be accurately modelled. There is an interface, based upon the MQTT messaging protocol, that allows input to and output from the world. The world we have created is mapped to the same location as the Anglova Scenario. We are able to use and extend the Anglova Scenario and it’s experimentation data within the world for conducting DAIS research.

Network defensive cyber operations (DCO) are inherently multi-domain, traversing different network segments and functional levels that encompass networking devices, protocols, services, applications and users. However, recent AI technologies threaten to complicate DCO as they can learn and adapt novel cyber-attack decision strategies to defeat countermeasures. Specifically, Reinforcement and Deep Reinforcement Learning (RL/DRL) are AI technologies for sequential decision-making in complex environments that have exceeded human master level performance in several domains through their ability to navigate the enormous state spaces of these environments. To investigate the effectiveness of AI-empowered autonomous cyber attacks, this work presents a preliminary study of DRL algorithms in training red AI agents in multi-domain computer networks. Employing a cyber network attack environment in the OpenAI Gym, the agents are trained to automatically establish and optimize their attack decision strategy. Different DRL algorithms are tested to evaluate the effectiveness against a selected set of network, service and application configurations, and to compare their stability, robustness and generalization characteristics. The results illustrate the potential of DRL-based cyber agents for researching new schemes to support cyber offence and defence operations.

While software using artificial intelligence and machine learning (AI/ML) is pervasive in many areas of society today, the use of these technologies to diagnose and treat medical conditions is limited due to a number of challenges associated with the trustworthiness of the results. This may include the inability to fully explain how an algorithm works inherent to the black-box nature of the system. Additionally, AI/ML may create a potential for bias and artifacts that cannot be validated due to the same limitations. In a medical application, the lack of transparency in how the system operates may lead to a loss of trust by users. Bayesian approaches that use computational modeling to quantify the level of uncertainty in a given result may provide a path towards improved confidence and use. In this paper, evidence from studies in a range of medical applications is presented and discussed, showing how Bayesian approaches can help to foster trust. A retrospective study using a publicly available dataset explored the feasibility of creating predictive models for early intervention in a Type 1 diabetes population. Creating the perfect model was not the goal of the exercise, rather the study aimed to demonstrate how Bayesian methods could be used to identify areas of uncertainty during model development. Feature selection was based on analytical assessment of various patterns found in the data. Models were trained, validated, and tested, generating uncertainty estimates. A two-feature Gaussian Naïve Bayes (GNB) model, using the previous five minutes and ten minutes of blood glucose values, showed similar results for predictive accuracy as a threefeature model that included average change over the preceding 30 minutes. The two-feature model was selected because it allowed for a more easily understood visualization of uncertainty. The 2-feature GNB achieved an AUC = .94. The model showed good sensitivity for exceeding the < 180 mg/dl limit, obtaining threshold prediction = 89.8% and normal range prediction = 90.8%. The sensitivity was lower for the < 70 mg/dl limit, attaining a sensitivity = 77.5%. Posterior probabilities showed differing levels of uncertainty in the prediction of high and low out-of-range conditions. The model demonstrated the feasibility of providing robust parameter estimates. Bayesian machine learning approaches to model uncertainty may improve the transparency, explainability, and applicability of AI/ML in medical treatment, realizing the promise to improve patient safety and outcomes.

As methods and access to gene synthesis and genetic engineering have become more advanced, the fear that malicious viruses and bacteria will be designed with the express intention of causing harm to humans has received increased attention. In the event that such biological weapons are deployed, the security community needs tools to rapidly recognize the threat and identify responsible parties. Therefore, a key question is whether or not a biological threat is manmade. Currently, experts are capable of qualitatively assessing whether specific genetic sequences are natural or man-made, but few objective criteria exist for characterizing the degree to which a sequence has been engineered. Additionally, progress has recently been made on the task of attributing an engineered gene sequence to a lab-of-origin using machine learning. However, the task of analyzing naturally occurring genetic sequences so as to automatically detect outliers that may have been genetically engineered has received comparatively little attention. This work proposes a method for generating a dataset of natural and engineered sequences that can be used as an input for training machine learning classifiers to perform automatic detection of human engineering in gene sequence data.

Computer simulation experimentation examined the effectiveness of different Unmanned Aircraft System (UAS) swarm configurations for identification and localization of survivors after a natural disaster using the DroneLab application. Swarms differed in terms of total number of drones and ratio of entities programmed to perform one of three different “personalities”—Relay, Social, and Antisocial. Relay behavior puts a high priority on maintaining proximity to the centroid of the swarm while also maintaining a distance to closest neighbor drones equal to half of the maximum WiFi range. Antisocial drones prioritize an expanding behavior, increasing the spread of the swarm, while the Social behavior prioritizes a contractive behavior resulting in a tighter swarm formation. All drones performed a local waypoint-based search behavior while conducting a spiral-out search pattern upon detecting four or more survivors within a 10-meter radius. Swarm configurations with different ratios of these behaving entities were assessed for mission completion, defined as time to find 90% of the survivors. Mission completions were recorded for four simulation scenarios consisting of two terrains (urban/rural) with two different distributions of survivors (naturalistic/randomized). Ten replications of 98 different drone configurations were evaluated. Statistically significant differences between time to mission completion between the terrains, between the two distributions, and among the iterations were revealed. Qualitative comparisons revealed differences in configurations that performed the best in each terrain. A few configurations performed well in all four scenarios. Moreover, the minimum number of entities needed for well-performing swarms was indicated. The work demonstrates the utility of computer experimentation and statistical analyses for developing a framework for swarm design for operational effectiveness.

Today’s battlefield increasingly incorporates emerging technologies using artificial intelligence. These systems not only provide unparalleled speed and accuracy, but also allow for digital models to be developed and tested in simulation prior to deployment, reducing the time and cost of acquisition. This holds additional promise for wargaming modeling and simulation for understanding the impact of complex, multi-domain operations on future force efficacy and structure. However, current modeling and simulation environments are not designed for simulating decentralized, intelligent systems at scale. Cloud computing has revolutionized how we scale computational capability, but was not designed for complex, low latency interactions between independently reasoning entities. This motivates new methods for characterizing and mitigating complexity to meet operational and mission requirements. We outline the challenges and opportunities for modeling and simulating large-scale multi-agent systems and identify future research areas that should address these challenges. We recommend that investment be placed in holistically understanding scalability from a cost-benefit perspective, measuring the impact on requirements, developing improved tools for understanding the dimensions of scalability, and formalizing specifications of the scalability requirements met (or not met) by available systems. We propose that a framework for reasoning over and adjusting the fidelity of various models within a system of systems is needed to meet development and testing requirements. Formal methods can be used to understand the limits on scalability as a function of objectives (e.g. speed, convergence, performance) and constraints (e.g. cost, compute, and time), optimizing resources to develop and test interacting artificial intelligence systems at scale.

Current and future military operations rely on the use of multiple dissimilar platforms ranging from unmanned and manned ground, aerial, naval, and space vehicles, the aggregation of which leads to Multi-Domain Operations (MDOs). The individual vehicular-based platforms are characterized by the presence of three components: (1) power/energy generation, (2) energy storage, and (3) mechanical/electrical/chemical/thermal-based energy demands. Inadequate understanding of the energy consumption, generation, and demand in combination with the energy reserves may limit mission effectiveness and efficiency, which in turn can reduce the effectiveness and flexibility of military operations. To mitigate energy degradation and maximize mission capabilities, energy and exergy characterizations for each individual component, platform, and system must be established and used to develop operational intelligence. Using Artificial Intelligence (AI) and Machine Learning (ML), energy and exergy flow characterizations may be developed and exploited. Using MATLAB/Simulink, vehicular models have been developed in combination with prediction-based algorithms using Artificial Neural Networks (ANNs) and Long-Short-Term-Memory (LSTMs) networks. The neural networks are used to; (1) approximate energy flow characterizations from on-board sensors which link dissimilar energy flows (electrical to thermal, or electricalmechanical to thermal, etc.), and (2) predict future energy flow characterizations which could be incorporated into a Model Predictive Control (MPC) Energy Management Strategy (EMS) in which energy degradation and efficiency can be altered, flexibly managing mission capabilities while improving resilience and survivability. The full paper will include a more thorough analysis of the individual predictive algorithms and the MPC-based control results.

Within computer vision, deep neural networks (DNNs) have gained tremendous popularity in recent years due to their ability to extract and classify visual features. As this technology has become more widespread, some of the shortcomings of the DNNs have become apparent. Recently, researchers have found that DNNs prefer to learn texture from visual signals rather than shape and have observed that more shape extraction correlates to better DNN performance. DNN have been applied to a vast number of problems, with excellent results in many of them, including object detection. The combination of DNN feature extractors with regression and region proposal techniques, like the Faster Region-Proposal Convolutional Neural Network (Faster R-CNN) and Single Shot Detector (SSD), have yielded promising results. Meanwhile, from the field of digital image processing, grayscale morphology extracts shape information using morphological operations. The Differential Morphological Profile (DMP) performs morphological opening and closings with varied structuring element sizes and computes the absolute difference between the resulting steps. The DMP provides a mechanism to improve shape extraction within DNN, and increase model robustness. To that end, a DMP-based neural network, DMPNet, has been created to assist DNN with extracting shape information by adding layers that perform DMP prior to the first convolutional layer. We use the DMPNet as a feature extractor for Faster R-CNN and SSD, and apply it to Maneuverability Hazard Detection in unmanned aerial system (UAS) imagery. The benefits of this approach include better explainability, lower training times, and models more tuned to shape information.

Object detection is increasingly used onboard Unmanned Aerial Vehicles (UAV) for various applications; however, the machine learning (ML) models for UAV-based detection are often validated using data curated for tasks unrelated to the UAV application. This is a concern because training neural networks on large-scale benchmarks have shown excellent capability in generic object detection tasks, yet conventional training approaches can lead to large inference errors for UAV-based images. Such errors arise due to differences in imaging conditions between images from UAVs and images in training. To overcome this problem, we characterize boundary conditions of ML models, beyond which the models exhibit rapid degradation in detection accuracy. Our work is focused on understanding the impact of different UAV-based imaging conditions on detection performance by using synthetic data generated using a game engine. Properties of the game engine are exploited to populate the synthetic datasets with realistic and annotated images. Specifically, it enables the fine control of various parameters, such as camera position, view angle, illumination conditions, and object pose. Using the synthetic datasets, we analyze detection accuracy in different imaging conditions as a function of the above parameters. We use three well-known neural network models with different model complexity in our work. In our experiment, we observe and quantify the following: 1) how detection accuracy drops as the camera moves toward the nadir-view region; 2) how detection accuracy varies depending on different object poses, and 3) the degree to which the robustness of the models changes as illumination conditions vary.

Object identification or classification has found many applications, ranging from civilian to defense application scenarios and there is a rising need for a both effective and efficient identification approach. One of the common methods for this task is using a neural network. However, it could be very difficult for such a network to obtain accurate answers due to complex environments, especially when the data is of single modality. In this paper, we attempt to build a combined deep learning model which takes two distinct data modalities to help us achieve high accuracy multimodal classification systems. An experiment is conducted on Multimodal Unmanned Aerial Vehicle Dataset for Low Altitude Traffic Surveillance (AU-AIR) using both visual and sensor data. We compare our results between a model trained with visual data only and another combined model trained with both visual and sensor data. Improved object classification performance is observed when the multimodal method is applied.

Historically, the development of armor ceramics can be largely described as heuristic. Recently, advanced machine learning algorithms are being developed to accelerate advanced material discovery. As many important material properties depend on microstructure, segmentation algorithms applied to scanning electron microscope (SEM) images enable quantification of identified features. The desired goal is to relate key image metrics to quantified physical properties and make useful performance predictions and improvements faster than otherwise possible. Collecting large image datasets with high signal to noise ratio, even with automation, can be laborious. Moreover, traditional methods of establishing image ground truth often rely on supervised hand-tracing, which precludes application to 1000’s of images. This study creates an approximate ground truth automatically using Otsu’s algorithm to evaluate large image data sets with varying signal to noise ratio and understand the influence of noise on network model efficiency. The robustness of a U-net algorithm, commonly used for image segmentation, was optimized by introducing artificial noise to the training data. Initial work assessed the performance of generative adversarial networks in applying artificial noise to the images. Next, a U-net was generated while incorporating artificial and real noise into the training and validation sets respectively. The impact of training using artificial noise upon the accuracy of the resulting U-Net in segmenting real images of low quality are described below.

Jamming, whether intentional or not, threatens stable wireless communications by impeding a transmitted signal. New jamming technologies are regularly developed and deployed for use, which behave differently from their predecessors in order to bypass defense mechanisms. This makes existing jammer classifiers difficult to implement in fast changing dynamic environments, since human intervention is needed every time a new jamming technology is introduced. Improper maintenance will result in misclassification of the technology or allow jammers to pass through defenses. These scenarios will greatly reduce the performance of wireless networks and increase the response time for recovering from these attacks. As 5G continues to become more widespread, and other faster networks are released, wireless data rates will continue to grow. This furthers the need for a faster and more reliable jammer classifier, as shorter interruptions in service will cause even more data loss to occur. Incremental learning (IL) is a technique in machine learning that allows the introduction of new information to a previously trained network. Using IL, it is possible to create classifiers that can grow in number of classes without the need to retrain a new network from nothing. This allows remote devices to learn to adapt in dynamic environments with far lower memory cost. In this paper, we developed an IL-based jammer classifier using software defined radio (SDR) to detect when a jammer is present and classify the type and learn to classify new technologies when the type has not been encountered before.

Advancement in remote sensing capabilities have led to unprecedented quantity and quality of data across a number of sensing modalities. It is now possible to outfit nearly any mobile platform not only with high resolution cameras, but also with inexpensive infrared and LIDAR sensors. For the specific goal of providing a comprehensive assessment of vehicle maneuverability, we address the problem of co-registering multiple sensor phenomenologies, such as visual, infrared, and LIDAR imagery collected from vehicle-mounted sensors. We show that a data fusion across these sensors provides invaluable information in hazard detection, localization, and classification. In addition, the co-registered measurements lead to the feasibility of enhanced heterogeneous data machine learning methods. This approach is verified on a dataset collected by the U.S. Army ERDC.

This paper introduces a novel distributed architecture for the distributed control of autonomous agents, with guarantees for optimal peak-to-peak disturbances attenuation. The resulted scheme is able to perform trajectory control while preserving a pre-specified (distance-based) formation of the autonomous double integrator agents with first order (linear) actuators. The controller designed method is based on a "disturbances as sets" approach while the scheme is completely distributed, by constraining the sensing and communications radius of each agent in the formation. Compensation of inter-agent communications induce time-delays are also discussed.

Limited expert time is a key bottleneck in medical imaging. Due to advances in image classification, AI can now serve as decision-support for medical experts, with the potential for great gains in radiologist productivity and, by extension, public health. However, these gains are contingent on building and maintaining experts’ trust in the AI agents. Explainable AI may build such trust by helping medical experts to understand the AI decision processes behind diagnostic judgements. Here we introduce and evaluate explanations based on Bayesian Teaching, a formal account of explanation rooted in the cognitive science of human learning. We find that medical experts exposed to explanations generated by Bayesian Teaching successfully predict the AI’s diagnostic decisions and are more likely to certify the AI for cases when the AI is correct than when it is wrong, indicating appropriate trust. These results show that Explainable AI can be used to support human-AI collaboration in medical imaging.

Much work has been dedicated to the exploration of Multi-Agent Reinforcement Learning (MARL) paradigms implementing a centralized learning with decentralized execution approach to improve human-like collaboration in cooperative tasks. Here, we introduce variations of centralized training to describe cases where shared/independent reward structure is utilized to improve learning by training agents in an intelligent way, and to analyze cooperative behavior in multi-agent systems. This work discusses implications for providing a classification of the recent MARL algorithmic approaches on the basis of their information sharing mechanism (e.g., reward, gradient, action, parameter, observation/state space sharing) on cooperative behavior in multi-agent systems.

Artificial Intelligence/Machine Learning (AI/ML) services implemented at the tactical edge on multiple, distributed low power sensors can take advantage of Cloud IoT services and processes to learn in complex data environments supporting evolving mission tasks and continuous improvement of algorithms through Cloud automation and management. Remote sensors and their associated functionality must exhibit resilience against adversaries and deceptive techniques, and operate securely in all domains. Amazon Web Services (AWS) IoT Greengrass uses machine learning models that are built in the cloud and deployed locally on remote sensors and IoT devices. Limited datasets can be used to train models and be refined as more data is available. AWS SageMaker can be used for scene detection using Images and other signals of interest resulting in alerts and notifications from deployed sensors. The quality of machine learning models can be improved through captured data from IoT Greengrass being returned to the Cloud and processed by AWS SageMaker. Labeling of data can be streamlined through Cloud services such as Amazon SageMaker Ground Truth using auto-segment, automatic 3D cuboid snapping, and other automated labeling features. Using AWS Cloud security services in addition to AI/ML and IoT services provides additional advantages supporting identity and certificate management, sensor and data analytics, and intrusion detection and prevention. The number of low-cost sensors supporting Cloud-based IoT remote sensing using AI/ML algorithms at the edge continues to grow. Raspberry PI, NVIDIA Jetson Nano and others, equipped with appropriate sensor devices, are readily available for building a distributed remote sensing network and communicating with the Cloud. This paper provides a detailed description of components and the initial results of building a small distributed remote sensing network using NVIDIA Jetson Nano edge devices equipped with inexpensive acoustic and image sensors using IoT Greengrass and AWS SageMaker for detecting and identifying several target types. Additional Cloud services were used supporting monitoring and auditing, ensuring a secure and resilient operating environment.

An awareness of activities in operational environments is key to the U.S. Army’s strategy and many sensors and spectral regimes are employed to this end. Advances in wide-spectrum acoustic sensors and compact high performance computational hardware have created opportunities for enhancing awareness. The Engineer Research and Development Center (ERDC) is researching infrasound sensing as a means of persistent, remote monitoring to provide battlefield awareness. Machine learning techniques are used to identify unique signatures in the battlespace’s infrasonic environment. Given the limited number of labeled data sets, unsupervised Gaussian Mixture Modeling (GMM) is applied to identify these signatures utilizing the Short-term Fourier transform (STFT) and resulting Power Spectrum Density (PSD). This study describes the process of sorting collected infrasound data into categories based on PSDs for application to GMM algorithms that identify a characteristic class labeling. Labels in relatively short time frames are then associated with features seen throughout a 24 hour cycle to produce synthetic samples. Several Support Vector Machines are trained and used to separate in-class verses outlier features in time segments of new data. Outlier counts exceeding a threshold, typically 50%, label new data segments as novel and subject to further processing. Finally, efforts are d escribed f or directional focusing the array using multiple elements/sensors to localize signatures or to emphasize the signatures from different directions. GPU accelerations will be applied wherever possible to improve local bandwidth and throughput.

The Inter-Domain Path Computation problem under Node-defined Domain Uniqueness constraint (IDPC-NDU) is a recently investigated topic for finding the effective routing paths on the multi-domain network topology as well as transportation. The objective of the IDPC-NDU is to find the shortest path in the multi-domain directed graph that traverses every domain at most once. Since the IDPC-NDU belongs to NP-Hard class, this paper proposes a novel two-level approach based on an Evolutionary Algorithm (EA) to solve it. The first level aims to determine the sequence of crossed domains using an improved Genetic Algorithm (GA), while the second one aims to locate the minimally costly path between two nodes among the entire domains. Furthermore, we devise an approach to represent a chromosome, which reduces the chromosome length to the number of domains. Experiments on numerous sets of instances were implemented to show the effectiveness and characteristics of the proposed algorithm.

The recent advances in machine learning (ML) and Artificial Intelligence (AI) have resulted in widespread application of data-driven learning algorithms. Rapid growth of AI/ML and their penetration within a plethora of civilian and military applications, while successful, has also opened new vulnerabilities. It is now clear that ML algorithms for AI systems are viable targets for malicious attacks. Therefore, there is a pressing need for better understanding of adversarial attacks against ML models, in order to secure them against such malicious attacks. In this paper, we present a survey of adversarial machine learning and some associated countermeasures. We also present a taxonomy of ML/AI system attacks that follow the same properties and characteristics, allowing them to be linked with different defensive approaches. A taxonomy is given for both attack and defense, and attacks proposed in the literature are categorized according to our taxonomy.

In this paper we explore the problem of open set object recognition in maritime domains. In recent years, deep neural-networks have gained popularity in ship detection and classification problems, but there is little related work in applying Out-of-Distribution (OOD) methods in the maritime domain, or handling vessel classes that fall outside the training data lexicon. One major issue is the open set recognition problem of detecting unknown ships or vessels in the wild. If little or no training data exists for a rare vessel class, what level of performance can we expect from a network when it encounters these objects at inference time? In related object-classification work, deep neural networks have been shown to incorrectly classify OOD samples with high confidence. We apply OOD detection methods to synthetic overhead imagery and a deep maritime-vessel classifier to benchmark performance during rare vessel encounters, and understand how to gracefully resolve these events.

In this paper, we present an end-to-end pipeline for deep learning applied to Automatic Modulation Classification (AMC). We begin by utilizing Modulation Classification Network (MCNET), a recently published cost-efficient convolutional neural network (CNN) with skip connections. Model efficacy is confirmed and the algorithm is advanced with hyper parameter and regularization adjustments, transfer learned with an augmented over-the-air data set, and then a computationally superior version is deployed to an edge device. The model is initially trained with the well-known 2018 DEEPSIG data set that includes 24 modulation schemes. Transfer learning utilizes the Experiments, Scenarios, Concept of Operations, and Prototype Engineering (ESCAPE) data set. The edge node device utilized, but is not limited to, an NVIDIA Jetson AGX XAVIER. Under ideal conditions, classification at the edge node resulted in 96% accuracy with 11 over-the-air modulation schemes. Inferences at the edge were up to 13 times faster than the non-optimized model.

SAGE (Sentry Agents), is a dynamic multi-agent-based framework for system automation. It is used for creation of virtual agents with defined behaviors and states. The SAGE framework provides the flexibility for agents to exist independently and also to interact and collaborate with each other for task execution. As part of the capabilities within SAGE, simulations can be generated. For this work, SAGE has been used to create scenarios for simulation of decision tasks with uncertainty. Decision-making is a challenging task. The complexities of decision-making increases when the information that supports and forms decision comes with any uncertainty. Uncertainty can be represented in different ways however for this simulation, the Uncertainty of Information (UoI) concept will be utilized. This allows understanding of the various sources of uncertainty and their impact on decision making. One of the UoI algorithms implemented is the LRM version which is based on operations research. This version computes an UoI value that is used as an input to the decision tasks within the simulation. In this paper, we detail the construction of those scenarios and its integration into SAGE.

Current paradigms for collecting data to train fieldable computer vision (CV) algorithms are inefficient and expensive in terms of time and resources, and they have limited ability to adapt to changing target signatures. By leveraging opportunistic sensing of the operator’s natural behavior (e.g., firing a weapon, placing markers on a map), it is possible to triage the data for CV algorithms. When paired with an After Action Review (AAR) to confirm target presence and location, the resulting pipeline can efficiently update the CV for changing target signatures. Subjects (n = 10) participated in a simulated mission in which they were asked to mark video frames in which targets belonging to three classes first appeared, then perform a brief AAR in which they were asked to mark target locations in these frames. The CV (Detectron2) was initially trained on one target class, then retrained on instances of a novel class acquired by this pipeline. We report that retraining the CV on as few as 8 images resulted in good localization performance (AP₅₀ 85.6, AR 42.5) on unseen test images to this novel class. Ten minutes of retraining (600 iterations) was sufficient for good performance (AP₅₀ 58.3, AR 27.5). Data augmentation via random occlusions and apertures (‘bubbling’) boosted the training set 192-fold and improved the ratio of hits to false alarms and improved resilience to naturalistic occlusion and small sizes (AP 80-100 and AR 80). These results support our approach as an efficient method to adapt CV via partially labeled operational data.

Unmanned aerial systems (UAS) equipped with visual sensors can be quickly deployed to map novel regions, a useful ability in GPS-denied regions, search and rescue operations, disaster response, and defense. Assisted by such a UAS, a ground vehicle could safely navigate a given region, aware of the potential hazards seen from airborne sensors. Here, we propose a pipeline for identifying and mapping maneuverable regions and objects pertinent to safe navigation (cars, barriers, etc.) in sequential imagery captured from UAS sensors. First we use a semantic labeling deep neural network for identifying roads, an object detection neural network for detecting hazards of known classes, and a model that uses linear features to detect potential road hazards in labeled road pixels. This visual evidence regarding maneuverability is collected across temporally-sequential images and is spatially fused into a single map via visual feature correspondence. Fusion of road evidence is done on a per-pixel basis while clustering techniques are used to find objects given a set of co-mapped detections. We show the use of this pipeline for the quick and automated creation of maps that contain useful information with regard to safe navigation of a region captured by UAS sensors. These techniques serve as a part in the development of a model for safe, ecient navigation of GPS-denied/rapidly changing regions through the use of UAS-enabled mapping.

DevOps engineers take many factors into account when assessing the suitability of AI/ML algorithms for operations. They rely on documentation about data requirements, parameter settings, theoretical limits, and operating characteristics to anticipate how well an algorithm will perform in the field prior to writing a single line of code. However, the USAF currently lacks code quality and documentation standards for AI/ML, which forces downstream consumers to make assumptions about the behavior of algorithms that often lead to unexpected results and wasted efforts. Therefore, we present a preliminary set of criteria by which to judge the maturity of word embedding algorithms in terms of reproducibility, testability, and documentation quality. We hope the criteria will grow into a set of quality requirements that govern how the DoD procures AI/ML capabilities, and eventually motivate the need for an ML development maturity model. Our nascent evaluation criteria surfaced during our own struggles with trying to replicate and compare the performance between a well-known word embedding algorithm (Word2Vec) and a custom graph embedding algorithm (IRI2Vec), procured specifically for the purpose of connecting missions across ATOs. We walk through two case studies in which we applied our evaluation criteria to assess the maturity of both Word2Vec and IRI2Vec

The Joint Force’s ability to integrate data and systems at speed and scale is of utmost importance to maintain a competitive edge in the data-driven arena of 21st century warfare. The current methods employed by developers to find system integration points, however, are ad-hoc, predominantly manual, and too costly – consider the explosion in volume and heterogeneity of the underlying databases that support modern-day operations. In this work, we present a data archeology and alignment workbench, known as Semantic Lifting of Integrated Messages (SLIM), that allows developers to quickly discover and document commonalities among disparate data. Our novel workbench uses background knowledge to constrain the geometric orientation of unfamiliar data in ways that are meaningful to developers. To accomplish this, SLIM uses a variety of Machine Learning algorithms to unify and characterize the meaning of data, and a dashboard of linked visualizations to convey semantic commonalities from different perspectives and dimensions. Additionally, the workbench maintains a web of execution provenance that developers can use to trace and probe results in order to assess relevancy and meaningfulness. As a demonstration, we describe two studies that use our process to quickly find non-trivial commonalities among multiple datasets, including the Universal Command and Control Interface (UCI) and an ontology about the arrival and departure of vehicles.