This PDF file contains the front matter associated with SPIE Proceedings Volume 11751, including the Title Page, Copyright information, and Table of Contents.

Introduction to SPIE Defense and Commercial Sensing conference 11751: Disruptive Technologies in Information Sciences V

The Laboratory for Physical Sciences is a DOD lab performing research in quantum computing, novel computer architectures, high performance computing, brain-inspired systems for learning, and application of machine learning to cybersecurity problems. This talk will provide an overview of ongoing research efforts at the lab, and then will drill down into work applying machine learning and other techniques to the task of malware analysis. This includes development of classifiers to determine if a given file is malware, generation of features through static analysis, disassembly, decompilation and dynamic analysis, aides to the human malware reverse engineer, and automated signature generation for family identification.

Modern computing systems demonstrate strong propensity for unintended, emergent computations and the related unintended, emergent programming models that enable or amplify cyber-attacks. Computing mechanisms built for a particular purpose and with particular intended models of execution prove to be capable of executing unintended computing tasks outside of their original specification and their designers and programmers' mental models. We start examining systems for signs of emergent behavior (with methods such as fuzz-testing) only after they are fully built. However, a system's exploitability models and propensity for emergent execution arise-and can also therefore be mitigated-already at the design stage.

The DNA molecule can be modeled as a quantum logic processor, and this model has been supported by pilot research that experimentally demonstrated non-local communication between cells in separated cell cultures. This modeling and pilot research have important implications for information sciences, providing a potential architecture for quantum computing that operates at room temperature and is scalable to millions of qubits, and including the potential for an entanglement communication system based upon the quantum DNA architecture. Such a system could be used to provide non-local quantum key distribution that could not be blocked by any shielding or water depth, would be simultaneous over any distance, and could not be electromagnetically interfered with or eavesdropped upon. The quantum DNA model also has implications for artificial neural networks and can provide architecture for a system of quantum random number generation.

The goal of this effort is to train Deep Learning (DL) models using synthetic Orthogonal Frequency-Division Multiplexing (OFDM) datasets to predict the modulation schemes of real OFDM signals without transfer learning. To facilitate our study, we generated a synthetic dataset, OFDM-O, that consists of 480k instances across four different modulations which include BP SK, QP SK, QAM16, and QAM64. Each instance with 16 OFDM symbols consists of 1280 IQ symbols. Since real OFDM instances have lengths of [2, 5, 44] OFDM symbols, the DL models are trained using short instances in order to overcome the instance length mismatch. Two datasets generated dynamically during training, OFDM-ro and OFDM-riq, are derived from dataset OFDM-O, by randomly choosing 5 consecutive OFDM symbols or 400 consecutive IQ symbols from each instance in OFDM-O at each epoch. 1-D Residual Neural Network (ResNet) models trained using three datasets achieve overall accuracies of 97.8%, 84.5% and 77.6% for OFDM-O, OFDM-ro and OFDM-riq, respectively. Cross validation of the three datasets shows that the ResNet model trained using OFDM-riq predicts the validation datasets of OFDM-O and OFDM-ro with high accuracy. Furthermore, a two-step validation is proposed during training of DL models where DL models are first validated with a synthetic validation dataset and then validated with real OFDM instances. Including a validation set with real signal allows us to terminate training before the DL model is over fit to the synthetic signals. The ResNet model trained using OFDM-riq correctly predicts 5 out of 7 short instances and all 5 long instances in the testing dataset of real signals. Both mis-classifications come from short instances of 2 OFDM symbols. Overall, the ResNet model trained using OFDM-riq can successfully predict the modulation schemes of real OFDM signals with high accuracy.

Hyperdimensional computing (HDC) is a type of machine learning algorithm but is not based on the ubiquitous artificial neural network (ANN) paradigm. Instead of neurons and synapses, HDC implements online learning via very large vectors manipulated to represent correlations among the various vectors, measured by a similarity metric. Yet this approach readily affords one-shot learning, transfer learning, and native error correction, which are standing challenges for traditional ANNs. Further, implementations using binary vectors {0,1} are particularly attractive for size, weight, and power (SWaP) constrained systems, particularly disposable robotics. The paper is the first to identify and formalize a method to completely clone trained hyperdimensional behavior vectors. Using shift maps, d-1 unique clones can be made from a parent vector of length d. Additionally, expeditionary robots with extraneous sensors were trained via HDC to solve a maze even when up to 75% of the sensors fed irrelevant data to the robot. Lastly, we demonstrated the resiliency of this encoding method to random bit flips and how different network topologies contribute to dynamic reprogramming of HDC robots. HDC is presented here though not to replace ANNs but to encourage integration of these complementary ML paradigms.

In both military and commercial domains, tasks are increasingly entrusted to autonomous systems and robots. These artificial intelligence (AI) systems are expected to be safe and intelligent, adapt to changing environments, and interact with other actors, both automated and human. In this paper, we present a framework and corresponding analytics for developing AI agents that possess (1) cognitive skills, including the ability to perform counter-factual reasoning and self-assessment and to exhibit human-like curiosity, biases, and errors; (2) the ability to learn complex tasks quickly with limited feedback; (3) the ability to coordinate and co-learn with human or AI teammates; and (4) the ability to function well over long time horizons (e.g. hours or days). Our framework is based on the theory of adaptive behavior called active inference, which was developed in computational neuroscience and psychology. Together with learnable deep factorized representations, the active inference provides the objective function, high-capacity predictions, and scalable computational mechanisms that enable AI agents to execute four processes fundamental to human cognition: learning, perception, planning, and simulation. We demonstrate the advantages of our AI solution in the domain of planning multi-agent maneuvers for executing area control missions. Our model achieves faster learning compared to the reinforcement learning baseline, producing faster point accumulation and game win rate.

With the advent of neural networks, users at the tactical edge have started experimenting with AI enabled intelligent mission applications. Autonomy stacks have been proposed for the tactical environments for sensing, reasoning and computing the situational awareness to provide the human in the loop actionable intelligence in mission time. Tactical edge computing platforms must employ small-form-factor modules for compute, storage, and networking functions that conform to strict size, weight, and power constraints (SWaP). Many of the neural network models proposed for the tactical AI stack are computationally complex and may not be deployable without modifications. In this paper we discuss deep neural network optimization approaches for resource constrained tactical unmanned ground vehicles.

Normalizations used in model reduction can be chosen to emphasize anything from computation reduction to parameter reduction. Choosing a normalization that emphasizes a model with a small number of parameters is useful when deploying a model onto machines with a limited communication rate, while choosing a normalization that emphasizes a model with a small computational cost is useful when deploying a model onto a machine for real-time sensor analysis. As such, we explore the effect of various normalizations used to prune kernel parameters on models trained on the ImageNet database.

Determining position within an indoor environment can be difficult when GPS signals become too weak. For this reason, alternatives are desired for indoor positioning systems (IPSes). The Bluetooth Low Energy (BLE) protocol is one alternative solution for IPSes. BLE is a low power wireless technology used for connecting devices with each other. There are two different methods for using BLE for localization: deterministic, and machine learning (ML) models. Each method uses a measured received signal strength indicator (RSSI) to determine distances from fixed, known locations. Deterministic models rely on empirical equations relating signal strength to distance, while ML uses collected signal strengths, or fingerprints, to learn positions. This paper assesses the robustness of an IPS system we built that uses BLE and ML by executing a distance fraud attack. A distance fraud attack causes intentional miscalculations of positions. The attack executed on the system assumes the attacker has network access and has compromised some small fraction of the receiving nodes. The results show a significant difference between the calculated positions of the system operating under benign conditions and operating under attack. We explore one possible defense against this attack by training an ML system for attack identification.

Image classification is a common step in image recognition for machine learning in overhead applications. When applying popular model architectures like MobileNetV2, known vulnerabilities expose the model to counter-attacks, either mislabeling a known class or altering box location. This work proposes an automated approach to defend these models. We evaluate the use of multi-spectral images to combat adversarial attacks. The original contribution demonstrates the attack, proposes a remedy, and automates some key outcomes for protecting the model’s predictions against adversaries. Similar to defending cyber-networks, we combine techniques from both offensive (“red team”) and defensive (“blue team”) approaches, thus generating a hybrid protective outcome (“green team”). For machine learning, we demonstrate these methods with 3-color channels plus infrared. The outcome uncovers vulnerabilities and corrects them with supplemental data inputs commonly found in overhead cases particularly.

Execution context isolation is a key security requirement in personal computers, edge devices, and more importantly on multitenant computing environments such as the Cloud. It is vital that data belonging to one context (e.g., a process, enclave, or virtual machine) cannot be accessed or modified by another context without explicit permission, particularly in consideration of a remote adversary. However, the level of context isolation provided by today’s systems is not well aligned with the security needs of personal users, cloud providers, and customers of cloud computing services alike. Current hardware enforces isolation at the architectural level. However recent high-profile attacks demonstrated that isolation guarantees are weak at the microarchitectural-level. To make matters worse, a lot of the defenses against microarchitectural defenses aim to protect specific side-channels rather than providing a more comprehensive solution. In this paper we describe our recent and ongoing efforts in providing holistic defenses against microarchitectural attacks.

Zigbee is a popular specification for Internet of Things (IoT) mesh networking that provides a suite of protocols built on the IEEE 802.15.4 standard for radio communication. The Zigbee protocol stack is designed as series of layers each with a specific set of functions for communicating data throughout the network. These protocols provide a comprehensive functionality for performing various network tasks, including commissioning new networks and devices, performing broadcasting, unicasting, groupcasting with end-to-end acknowledgement, securing network traffic through AES-128 encryption, and full-packet message authentication. Security features of the Zigbee protocol alone may not be a complete solution for deploying secure IoT networks. It has some vulnerabilities and real-world attacks as discussed in this paper. Zigbee may be improved upon or added to for the purpose of securing it using real-time anomaly detection in IoT Local Area Networks (LANs).

Autonomous systems will operate in highly contested environments in which it must be assumed that adversaries are equally capable, agile and informed. To achieve and sustain dominant performance in such environments, autonomous systems must be able to adapt through online machine learning while managing and tolerating attrition - that is, improve their performance quickly, even over the duration of a single engagement with principled asset losses. However, there are novel challenges to adapting effectively in such environments. We present an approach that leverages several recent innovations in reinforcement learning, distributed computing and trusted consensus algorithms such as Blockchain. We note that multi-agent systems operating in contested environments must leverage their redundancy for learning while also remaining resilient with respect to component failures and com- promises. In particular, to enable and accelerate learning, such systems will have to allow some number of components to operate sub-optimally to achieve the right exploration-exploitation balance needed for rapid and effective learning. At the same time that some number of components are possibly being sacrificed due to sub-optimal performance, the underlying mission of the system must be maintained. This leads to challenges in distributed trusted computing such as Byzantine agreement problems. Simulations demonstrating these various tradeoffs using epidemiological models are presented.

Leading organizations review concepts and principles of safeguards risk management systems for implementing, verifying and measuring processes and procedures in large, distributed enterprise organizations. In light of potentially disruptive technologies, this Qualitative Risk Assessment methodology evaluates and quantifies risk exposures associated with operations, finances, regulation, organizational, reputational, and technology. Risk assessments should be conducted as soon as they can, to mitigate potential disruption. Based on assessment results, organizations will be better able to gauge vulnerabilities and target areas for prioritizing risks. Due to the rate of technological advances and adaption, it’s critical that risk assessments be conducted continuously.

Iris recognition is one of the most accurate biometric recognition techniques, however off-angle iris recognition has yet to have an established comprehensive recognition framework. This is due to the difficulties in the recognition of off-angle iris image inconsistencies within the iris patterns when gaze deviations are present. In this work, we investigate different iris normalization techniques and compare their performance. The two methods under investigation include elliptical normalization and circular normalization after frontal projection of off-angle iris recognition. Elliptical normalization samples the iris texture using elliptical segmentation parameters: 𝑥, 𝑦, 𝑟₁ , 𝑟₂ , θ where 𝑥, 𝑦 are coordinates, 𝑟₁, 𝑟₂ are the radius, and θ is the orientation. Also, when investigating circular unwrapping, we will be using the ellipse segmentation parameters to estimate the gaze deviation. The image will be projected back to a frontal view using perspective transformation. Then, we segment the transformed image and normalize using the circular parameters: 𝑥, 𝑦, 𝑟 where 𝑥, 𝑦 are coordinates and r is the radius. We further investigate if: (i) elliptical normalization or circular unwrapping recognition performance is higher, and (ii) if the two segmentation methods in circular unwrapping increase the recognition efficiency