This PDF file contains the front matter associated with SPIE Proceedings Volume 11419, including the Title Page, Copyright information, and Table of Contents.

For realistic synthetic imagery, radiative transfer methods coupled with large mesh geometry provide the most scientifically accurate way to model a scene. Radiative models typically use ray-tracing techniques to determine where radiative energy is coming from or moving to. This work presents an approach to making a ray query Geometry Engine that actively stores large-scale, terabyte-sized geometry in out-of-core memory on parallel general purpose processors. Procedures for geometry distribution, structures for efficient ray-tracing, and the ray query API are discussed. Geometry distribution uses Morton codes with parallel sorting routines to create geometry scene-chunks that are distributed among processing nodes. Each scene chunk is then broken down using a bounding volume hierarchy (BVH) using axis-aligned bounding boxes (AABB). The BVH allows for efficient ray tracing of the geometry. The ray query engine API allows client-side programs, such as sensor models and radiative transfer models, which exist on the same high performance computer to efficiently identify intersected geometry given directed rays and collect individual geometry elements. The geometry has key values that can uniquely identify data from solver programs. Scalability, partition timing, and ray timing results are presented.

Diabetic Retinopathy (DR) is one of the most common eye diseases related to diabetics. If the diagnosis and treatment are conducted too late, it may result in various degrees of vision loss, even blindness. Therefore, individuals with diabetes should have a regular annual eye exam. Studies showed that early detection can prevent vision loss in earlier stages. However, in places such as undeveloped or developing countries, and even sometimes rural areas in developed countries, may not have enough resources for DR screening. Furthermore, even though these places may have adequate equipment, the diagnosis may take a few days to obtain results for the analysis of the ophthalmologists. Developing an automated detection algorithm is an emerging research area to diagnose DR remotely using a retina image. Localizing the optic disc and fovea is an essential task in these DR detection algorithms. After locating the optic disc, finding other components of the retina is easier. Technological developments in recent years enable the acceleration of diagnosis of such diseases including DR. Deep learning techniques are becoming an essential part of the medical field. In the last years, there have been many attempts to automatize the analysis of medical disorders such as breast cancer, glaucoma, diabetic macular edema, and diabetic retinopathy. In this paper, we presented the utilization of a pre-trained deep learning framework to localize the optic disc in the retina images. Using the transfer learning approach for AlexNet with a linear regression output, we localized the optic disc center. Retina images with labeled ground truth values of optic disc center were used to retrain the AlexNet. We tested our proposed deep learning-based optic disc localization approach with three different publicly available datasets including EyePACS, Messidor, and IDRID. Based on the results, the deep learning-based optic disc localization method shows high detection accuracy. The best results for optic disc detection were observed with cross dataset images as the accuracy of 88.35%, while a 97.66% testing accuracy was observed for the merged dataset using transfer learning approach for the pretrained AlexNet.

If electro-optic conversion of current photonic NNs could be postponed until the very end of the network, then the execution time is simply the photon time-of-flight delay. Here we discuss a first design and performance of an all-optical perceptron and feed-forward NN. Key is the dual-purpose foundry-approved heterogeneous integration of phase-change-materials resulting in a) volatile nonlinear activation function (threshold) realized with ps-short optical pulses resulting in a non-equilibrium variation of the materials permittivity, and b) thermo-optically writing a non-volatile optical multi-cell (5-bit) memory for the NN weights after being (offline) trained. Once trained, the weights only required a rare update, thus saving power. Performance wise, such an integrated all-optical NN is capable of < fJ/MAC using experimental demonstrated pump-probe [Waldecker et al, Nat. Mat. 2015] with a delay per perceptron being ~ps [Miscuglio et al. Opt.Mat.Exp. 2018] has a high cascadability.

One needs a good communication cost model to understand how best to optimize the off-loaded computation in a tactical environment. In practical scenarios, complications also arise from the in-band and out-of-band channel congestion interference. This can happen due to the intentional and/or unintentional adversary equipment and will affect both the nature of algorithms and the sequence of steps in their execution. As a first step towards solution of the above problems, we present a model for a multi-node tactical network with resource constraints with and without presence of some adversary nodes.

During multi-robot simultaneous localization and mapping (SLAM) tasks, a team of robot agents must efficiently synthesize a global map from individual local maps. One method to accomplish this is reserving a subset of robots to perform data fusion and assigning the others to collect and generate local maps. While this solves, the division of labor problem, it requires humans to explicitly designate which robots handle what tasks and it does not scale well to large teams. Moreover, when robots are operating in tactical environments, they may be placed on a heterogeneous team, with limited or unreliable communication infrastructure. To assist with the task of role assignment, we describe a novel decentralized message passing algorithm, for what we are calling Distributed Leader Consensus (DLC). DLC helps a set of agents self-organize into structured groups by giving them the ability to autonomously come to a consensus on the group leader. Our approach is entirely distributed, easily configurable, and is robust to agents being dynamically added to or removed from the system. DLC may be configured to limit group sizes, assign multiple leaders, and select leaders based on computational power and/or physical proximity. We test our approach in simulation by having a set of agents adapt to changing teammate availability.

Here we compare the latency performance of active and passive offloading decisions pathways which we use with our adaptive computing engine. The passive pathway will query nodes for offloading after a job has been submitted, while the active pathway will query system information before jobs are submitted to create an offloading plan. The offloading plan reduces latency and communication overhead by cutting out the offloading query required for each job under the passive pathway. Overall we see a 6ms reduction in latency when using the active pathway in a multi-node environment connected over Wi-Fi.

US Military need to know more and sooner (situational awareness) to allow rapid, decisive action on the emerging multi- domain battlefield with multi-faceted internet of battlefield things, big data, AI, and tactical cloud. Now, and even more so in the future, the battlespace is characterized by highly distributed processing, heterogeneous and mobile assets with limited battery life, communications- dominated but restricted network capacity, and operating with time-critical needs in a rapidly changing hostile environment. Distributed and low power edge processing is one of the essential technologies for maintaining overmatch in various emerging operational and contested environments for various technical groups across DoD COI communities including C4I, microelectronics, cyber, autonomy, human systems, Electronic warfare, and Weapon systems. Number innovations in computing hardware, algorithms, software, networking and complex Military applications at the edge are helping DoD to take advantage of this new paradigm of computing at the edge. This talk will discuss recent advances in trusted computing at the edge for Military applications.

Counterfeiting of integrated circuits (ICs) is a rising issue with a huge economic impact on the electronics industry. With the electronic component supply chain growing more complex due to globalization, and parts coming from a diverse set of suppliers, counterfeit electronics have become a major challenge. Given the critical engineering applications that use ICs consistently, such a growth in the supply chain’s complexity has led to serious security threats due to potential insertion of noncompliant hardware components. In this paper, we propose a game-theoretic framework to analyze the interactions between a hardware supplier (attacker), and the IC component testing site of a company (defender). We formulate the problem as a non-cooperative zero-sum game in which the attacker decides on the type of counterfeit component to insert into the chain while running the risk of detection and its collateral damages. Meanwhile, the defender, seeking an optimum utility payoff while contending with limited resources, must decide on the best testing and detection strategy. Simulation results assess the performance of the proposed game-theoretic framework and provide further insights into the effectiveness of our solution.

Quantum computing is coming; and, with it, the security of modern cryptography will be compromised. Technologies such as industrial control systems, banking, and smartphones depend on cryptography to ensure the confidentiality and integrity of operations. Any data sent over the internet relies on cryptography. In some cases, information is so sensitive that if the encryption is broken years in the future the results could be catastrophic. A post-quantum cryptographic solution is required today. Variable World Length is a proposed quantum-proof symmetric cryptography algorithm. It is low-labor, low-latency, and low-power; with a keyspace of 10⁵¹¹ in its simplest form (for comparison, AES-256 is 2²⁵⁶). Variable Word Length performs bit manipulation to send “words” of varied length. Without the key, an attacker would need to brute force all possible combinations. Due to the design of the algorithm, some of these combinations will decrypt real plaintext that was not the encrypted message. Additionally, the keyspace can be increased without an exponential tax on computing resources.

Performing feature extractions in convolution neural networks for deep-learning tasks is computational expensive in electronics. Fourier optics allows convolutional filtering via dot-product multiplication in the Fourier domain similar to the distributive law in mathematics. Here we experimentally demonstrate convolutional filtering exploiting massive parallelism (10^6 channels, 8-bit at 1kHz) of digital mirror display technology, thus enabling 250 TMAC/s. An FPGA-PCIe board controls the ‘weights’ and handles the data I/O, whereas a high-speed camera detects the inverse-Fourier transformed (2nd lens) data. Gen-1 processes with a total delay (including I/O) of ~1ms, while Gen-2 at 1-10ns leveraging integrated photonics at 10GHz and changing the front-end I/O to a joint-transform-correlator (JTC). These processors are suited for image/pattern recognition, super resolution for geolocalization, or real-time processing in autonomous vehicles or military decision making.

Classical optimization algorithms in machine learning often take a long time to compute when applied to a multi-dimensional problem and require a huge amount of CPU and GPU resource. Quantum parallelism has a potential to speed up machine learning algorithms. We describe a generic mathematical model to leverage quantum parallelism to speed-up machine learning algorithms. We also apply quantum machine learning and quantum parallelism to a 3-dimensional image that vary with time as well as tracking speed in object identification.

Autonomous systems, including self-driving cars and air vehicles, have caught the imagination of the press and the public. However, broader adoption of such systems in safety-critical applications has been the subject of intense debate and scrutiny. The stunning performance of deep learners compared to extant methods, including pattern matching, statistical methods, and legacy machine learning algorithms, has taken the research world by storm. This has naturally led the DoD community to ask the question: “How do we harness this technology being unleashed upon the world?” Before we answer this question, however, it is important to note that trust is integral to DoD applications, including autonomous systems, and ensuring reliable system operations is paramount. Therefore, we need strategies that harness deep learning algorithms to provide the DoD with autonomous systems that are robust, secure, timely and dependable.

Artificial Intelligence and Machine Learning (AI/ML) based solutions contain inherent vulnerabilities that are of growing concern as investment is made into applying these technologies to gain a strategic advantage in speed of response, analysis, and maneuver for both private industry and government applications. Current cybersecurity practices and tools must evolve to handle new vulnerabilities inherent to AI/ML enabled systems across all domains. This paper will explore how the tools and techniques to defend against attacks and exploits, both on AI/ML, and with AI/ML, fall short from the tools typically used by today’s cybersecurity professional. It will explore both the intrinsic vulnerabilities due to model failure points and data poisoning strategies as well as address concerns that arise when our adversaries use AI/ML tools to their advantage. Some of these challenges present themselves as very advanced strategies brought on by nation state actors who have both time and resources. Other threats to our AI/ML systems are much less sophisticated but still seem to slip through the cracks because AI/ML is only beginning to gain significant momentum in real world applications.

Machine Learning (ML) technologies have been widely used for blind signal classification problem (Modulation Classification) to automatically identify the modulation schemes of Radio Frequency (RF) signal from complex- valued IQ (In-phase Quadrature) samples. Traditional ML methods usually have two stages where the first stage is to manually extract the features of the IQ symbols by subject matter experts and the second stage is to feed the features to an ML algorithm (e.g., a support vector machine) to develop the classifier. The state- of-art technology is to apply Deep Learning (DL) technologies such as Convolutional Neural Network (CNN) or Recurrent Neural Network (RNN) directly to the complex-value IQ symbols to train a multi-class classifier. In this effort we are focused on the modulation and coding rate classification problems for multi-carrier Orthogonal Frequency-Division Multiplexing (OFDM) signals. The generated OFDM data set consists of over 480k instances across four different modulations and three different coding rates which include BPSK + 1=2, BPSK+3=4, QPSK+1=2, QPSK+3=4, QAM16+1=2, QAM16+3=4, QAM64+2=3 and QAM64+3=4. Deep Learning (DL) models can successfully catch the features of OFDM modulations and identify the modulation scheme from the baseband IQ symbols. Four DL models including Residual Neural Network 34(Resnet34), Resnet18, Squeezenet and Long Short-Term Memory (LSTM) trained over the data set can achieve over 94% overall accuracy for signals across 􀀀10dB to 10dB with step size of 4dB. Among them, Squeezenet and LSTM models have much smaller model sizes which can be easily loaded in resource-limited edge computing platforms. To the best of our knowledge, coding rate classification of OFDM signals has not been studied in previous works. Therefore, three different DL models including a single-stage coding rate classifier, a combination of modulation and coding rate classifier, and a two-stage coding rate classifier are developed to identify the coding rate. However, coding rates can only be identified for low-order modulation schemes BPSK and QPSK but not for high-order modulation schemes QAM16 and QAM64. Further investigation is needed to understand the coding rate classification for high-order QAM signals.

Today's sensors and analysis systems produce huge amounts of data and one of the main challenges is how to enable users to find relevant data in time. Software systems supporting queries must provide suitable filter mechanisms. Through this filtering, the full dataset can be reduced to a manageable amount of relevant data. Respective filter criteria can be chained together in the form of Boolean expressions using disjunctions and conjunctions, and the resulting hierarchical structures can be further simplified by parentheses. The practical use case that we present in this publication consists of a web application to access potentially large data sets using a defined set of metadata based catalogue entries. This application currently supports the specification of filter criteria, which can be concatenated by the user through a single global conjunction in a at hierarchy. The underlying query language supports more complex queries using any combination of conjunctions, disjunctions and brackets. There is a user requirement to extend the expressiveness of the client search queries so that the full scope of the query language can be leveraged meaningfully. One problem is how such complex search queries can be graphically structured and visualized in a clear and comprehensive way. There exist different approaches for the graphical visualization of program code and query languages. Some of these approaches also support the graphical editing of these representations by the user. Examples for such frameworks or tools are Blockly, Scratch or Node-RED. In this publication, an analysis of the applicability of such frameworks in the existing web application is presented. For this purpose, operational constraints and exclusion criteria that must be fulfilled for the use in our application are identified. This results in a selection of a framework for a future implementation.

Face in video recognition (FiVR) is widely used in video surveillance and video analytic. Various solutions have been proposed to improve the performance of face detection, frame selection and face recognition in FiVR systems. However, all these methods have a common inherent ceiling", which is defined by the source video's quality. One key factor causing face image quality loss is video compression standards. To address this challenge, in this paper, first, we analysis and quantify the effects of video compression on the FiVR performance; secondly, we propose to use deep learning based model to mitigate artifacts in compressed input video. We apply the image based convolutional auto-encoder (CAE) to extract the features of input face images and restore them towards less artifacts. From the experimental results, our approach can mitigate artifacts on face images extracted from compressed videos and improve the overall face recognition (FR) performance by as much as 50% in TPR (True Positive Rate) at the same FPR (False Positive Rate) value.

Mobile Ad hoc Network or MANET is a wireless network that allows communication between the nodes that are in range of each other and are self-configuring. The distributed administration and dynamic nature of MANET makes it vulnerable to many kind of security attacks. One such attack is Black hole attack which is a well known security threat. A node drops all packets which it should forward, by claiming that it has the shortest path to the destination. Intrusion Detection system identifies the unauthorized users in the system. An IDS collects and analyses audit data to detect unauthorized users of computer systems. This paper aims in identifying Black-Hole attack against AODV with Intrusion Detection System, to analyze the attack and find its countermeasure.

We present and evaluate the performance of our Network Link Outlier Factor (NLOF) for detecting transmission channel faults in communication networks. An NLOF is computed for each transmission channel in a network under management using the throughput values derived from ow data. Throughput values of flows are clustered in two stages, outlier values are determined within each of the clusters, and then ow outlier ratios determine the outlier score for each transmission channel (link). Specifically, we first cluster the throughput of flows into the set of clusters we believe will naturally exist in a network and then identify the outliers within those throughput clusters. Our technique to detect network transmission channel faults consists of: 1) ow throughput clustering, 2) ow throughput outlier detection using an outlier score, 3) tracing flows on the network topology using routing information, and 4) network link outlier score computation from ow outlier scores.

During the past decade, smartphone developers have been improving security mechanisms in smartphone runtimes out of necessity. Unfortunately, these security improvements have frequently been left unimplemented in the smartphone’s corresponding baseband controller, and notable exploitations have occurred as a result. Developers of all cyber physical systems should take note: any processor running software could be the target of an attack; especially if that software processes raw inputs from sensors. In this paper we investigate specific exploitations that have been proven effective against baseband controllers, and examine the methodologies recommended by the exploit authors to protect against them. We conclude by presenting some design and implementation techniques that can be used to enhance the security posture of mindful developers of cyber physical systems.

In this paper, we evaluate the performance of several flow features to classify the network application that produced the flow. Correlating network traffic to network applications can assist with the critical network management tasks of performance assessment and network utilization accounting. Specifically, in this work we evaluate three engineered flow features and three inherent flow features (number of bytes, number of packets, and duration). For engineered features, we evaluate three host communication behavior features proposed by the authors of BLINC. Our experiments uncover the classification power of all combinations of the three engineered features in conjunction with the three inherent features. We utilize supervised machine learning algorithms such as k-nearest neighbors and decision trees. We utilize confidence intervals to uncover statistically significant classification differences among the combinations of flow features.

Mobile ad hoc networks (MANETs) pose challenges distinct from commercial wireless networking, including a need for decentralized routing. We consider the barrage relay network (BRN) routing algorithm for unicast communications, beginning with a simple connected graph model (CGM) and then more physically motivated models. We show geometrically and through simulation that lower bound on the domain of utilized nodes nodes converges to an elliptical region in the high-density limit. In this bound, the number of utilized nodes scales with the (1.5)-power of the distance between source and destination. We conclude with recommendations on how to reduce load while maintaining reliability.