This PDF file contains the front matter associated with SPIE Proceedings Volume 12538, including the Title Page, Copyright information, Table of Contents and Conference Committee lists.

The National Science Foundation’s Directorate for Technology, Innovation and Partnerships (TIP) engages the nation’s diverse talent in strengthening and scaling the use-inspired and translational research that drives tomorrow’s technologies and solutions. TIP provides education, tools and funding to spur innovation, accelerate the development of breakthrough technologies and speed solutions forward. Learn more at beta.nsf.gov/tip/latest

In a 2022 AUSA paper, Lyons and Johnson write, "Based on a theory of complementarity and domain interconnectedness, MDO works by creating combinations of interacting effects—physical, nonphysical, mental and moral—to weaken and collapse the enemy’s will to resist... These effects demand integration at very different ranges and moments in time that account for rates of movement and changes in location and disposition." It is straightforward to see how modern AI systems can solve well-bounded problems, within single domains and even across domains. To give just a few examples, AI is making a difference in sensing, information fusion, and control for autonomous robots (physical movement on a small scale of time and space); in defending against cyber attacks (nonphysical actions across large-scale networks); in decision support for command and control (variable scales of time and space with abstraction and cross-domain combination). There are pitfalls in the application of AI to military challenges, however, and there remain open problems on which research as made relatively little progress over the decades. This talk provides an overview of what we may expect for future trends, the challenges we will likely face, and where we believe resources can be best allocated.

With the acceleration of Artificial Intelligence (AI) innovation in recent years, organizations across the public sector have been racing to adopt AI into their products and processes. With respect to National Security, AI adoption offers the potential to automate the operation of valuable assets and provide critical warfighter support across the observe–orient–decide–act (OODA) cycle. As we progress towards integration of AI-enabled mission critical systems, the robustness of these systems must be verified, especially against adversarial exploitation. We present a process for red-teaming AI systems used in real-world decision-making. We consider attacks on target AI systems in the context of the larger concept of operations (CONOPS) and use threat modeling to narrow down likely attack vectors.

Penetration testing is used by numerous organizations, including most in the defense sector, to validate their IT security; however, it has typically been a manually intensive process. Artificial intelligence presents a prospective solution to this challenge, with automated active sensing offering the potential to identify vulnerabilities that time-limited human penetration testers cannot. This paper goes beyond simple automation and presents and evaluates a system that brings together an explainable artificial intelligence technology, a network scanning and modeling technology and an attack automation technology to make penetration testing an optimization problem that can be solved by machine learning.

In this work we present a new concept we call self-assessment which leverages techniques from the self-supervision literature to estimate algorithm performance and certainty in real-time. Our self-assessment approach enables AI/ML systems to determine what they know, how well they know it, identify what is fundamentally knowable about a scene, and highlight confusing or un-reliable data for further investigation. This approach has applications to identifying out-of-domain data or flagging unexpected changes in operational contexts that can otherwise reduce AI/ML trustworthiness.

The capabilities of convolutional neural networks, and in fact all manner of artificial intelligence and machine learning capabilities, to explore data in various fields has been documented extensively throughout the literature. One common challenge with adopting AI/ML solutions, however, is the issue of trust. Decision makers are rightfully hesitant to take action based solely on “the computer said so” even if the computer has great confidence that it is correct. There is obvious value in a system that can answer the question of why it made a given prediction and back this up with specific evidence. Basic models like regression or nearest neighbors can support such answers but have significant limitations in real-world applications, and more capable models like neural networks are much too complex to interpret. We have developed a prototype system that combines convolutional neural networks with semantic representations of reasonableness. We use logic similar to how humans justify conclusions, breaking objects into smaller pieces that we trust a neural network to identify. Leveraging a suite of machine learning algorithms, the tool provides not merely an output “conclusion”, but a supporting string of evidence that humans can use to better understand the conclusion, as well as explore potential weaknesses in the AI/ML components (whether as a result of lack of sufficient training data, adversarial attempts to corrupt the system, etc.). We have applied this system to problems of object detection and semantic segmentation of images. This paper will provide an in-depth overview of the prototype and show some exemplar results.

One of the inputs to a flight planning system is human-generated Notice to Airmen (NOTAM) entries that are used to alert flight crews of potential hazards that may be encountered when flying a specific mission. The text descriptions contained within the NOTAM are heavily abbreviated domain-specific free text, not standardized, and can vary widely depending on how the issuer chooses to describe the situation. Automated flight planning systems (autoplanners) do not parse the text description to determine accuracy or viability. Instead, a 4-letter code (known as a Q-Code) is entered and recorded by the human issuer to aid in interpreting the NOTAM contents. If Q-Codes are improperly entered, wrong, incomplete, or not specific to the issue being described, NOTAMs may be erroneously interpreted and the autoplanner may generate sub-optimal flight plans. As a solution to the problem, we developed machine-learning based text classification models where the inputs are NOTAM text descriptions and the outputs are predicted Q-codes for each text description. Such a solution would make autoplanner systems more robust by verifying and correcting incorrect NOTAMs automatically.

A major concern for artificial reasoning systems to achieve robustness and trustworthiness is causal learning where better explanations are needed to support the underlying tasks. Explanations for observational datasets without ground truth presents a unique challenge. This paper aims to provide a new perspective on explainability and causality by combining the two together. We propose a model which extracts quantitative knowledge from observational data via treatment effect estimation to create better explanations through comparison and validation of the causal features with results from correlation-based feature relevance explanations. Average treatment effect (ATE) estimation is calculated to provide a quantitative comparison of the causal features to the relevant features from explainable AI (XAI). This approach provides a comprehensive approach to generate robust and trustworthy explanations via validations from both causality and XAI to ensure trustworthiness, fairness and bias detection within the data, as well as the AI/ML models for artificial reasoning systems.

Battles and simulated battles are typically analyzed using standard statistics, e.g. the number of units damaged, the duration of combat, or the adherence to battle plan. Higher-level descriptors, however, are needed to truly understand and describe a battle. For example, these standard statistics are incapable of detecting or describing collaboration. Our work implements an NLP-inspired algorithm which analyzes battle actions and detects higher-level similarity among battle actions. Additionally, our work suggests possible applications of this technique.

One of the major issues limiting the adoption of machine learning (ML) in applications where accuracy is critical is failure of an otherwise accurate system. Recent work has developed tools to independently measure the competency of machine learning models and the conditions that drive these competency differences. The purpose of this paper is to explore ways to detect and mitigate adversarial attacks by using the same tools to assess competency. We will introduce BAE System’s MindfuL software which assesses ML competency under varying environmental conditions. We then consider a few different types of adversarial attacks and describe detection experiments. We examine the predicted performance and strategy and use that information to detect adversarial attacks. We will present the results of these experiments and discuss the implications of this work and potential future directions for this research.

In this work, we aim to develop novel cybersecurity playbooks by exploiting dynamic reinforcement learning (RL) methods to close holes in the attack surface left open by the traditional signature-based approach to Defensive Cyber Operations (DCO). A useful first proof-of-concept is provided by the problem of training a scanning defense agent using RL; as a first line of defense, it is important to protect sensitive networks from network mapping tools. To address this challenge, we developed a hierarchical, Monte Carlo-based RL framework for the training of an autonomous agent which detects and reports the presence of Nmap scans in near real-time, efficiently and with near-perfect accuracy. Our algorithm is powered by a reduction of the state space given by a transformer, CLAPBAC, an anomaly detection tool which applies natural language processing to cybersecurity in a manner consistent with state-of-the-art. In a realistic scenario emulated in CyberVAN, our approach generates optimized playbooks for effective defense against malicious insiders inappropriately probing sensitive networks.

Malicious cyber actors are well-prepared and trained to negatively impact organizations, assets, or individuals. Training exercises help to mitigate these adverse effects by enabling cyber defense analysts to demonstrate their abilities to respond to cyber incidents in real-time. When a security incident occurs, the response team must quickly contain the situation, diagnose the problem, and return to a previous good state so the system down time is minimized, and the issue does not spread to other systems. Training exercises allow both red (attack) and blue (defense) teams to practice in a secure, isolated environment ensuring all team members have the necessary experience if and when incidents occur. This paper describes simulated cybersecurity research scenarios based on open-source data repositories of adversary tactics and techniques including published frameworks by MITRE and NIST. Artificial Intelligence/Machine Learning (AI/ML) techniques and methods supporting cybersecurity research will also be discussed with an emphasis towards cloud environment frameworks. Results from using AL/ML techniques, cloud-based tools and methods, cyber tabletop exercises for cybersecurity research scenarios and use cases will be included.

Data manipulation could alter the performance of joint all domain command and-control decisions (JADC2). We present a Bayesian decision theoretic approach for adversarial forecasting when the underlying data collected over time is subject to attack from intelligent adversaries. Proposed adversarial risk analysis-based framework allows incomplete information and uncertainty. We solve the adversary’s poisoning decision problem where he manipulates batch data being inputted into the forecasting method of statistical autoregressive models. The findings show the vulnerability of forecasting models under adversarial activities. We discuss potential defender strategies.

Detecting malicious activity using a network intrusion detection system (NIDS) is an ongoing battle for the cyber defender. Increasingly, cyber-attacks are sophisticated and occur rapidly, necessitating the use of machine/deep learning (ML/DL) techniques for network intrusion detection. Traditional ML/DL techniques for NIDS classifiers, however, are often unable to sufficiently find context-driven similarities between the various network flows and/or packet captures. In this work, we leverage graph representation learning (GRL) techniques to successfully detect adversarial intrusions by exploiting the graph structure of NIDS data to derive context awareness, as graphs are a universal language for describing entities and their relationships. We explore several methods for NIDS data graph representation at both the network flow and packet level utilizing the CIC-IDS2017 dataset. We leverage graph neural networks and graph embedding algorithms to create a context-aware network intrusion detection system. Results indicate that adding context derived from GRL improves performance for detecting attacks. Our highest-scoring classifier incorporated both GNN embeddings and flow-level features and achieved an accuracy of 99.9%. Adding GRL methods to augment the flow/packet features improved accuracy by as much as 52.41%.

This study conducts three experiments on adversary decision-making modeled as a graph. Each experiment has the overall goal to understand how to exploit an adversary’s decision-making in order to obtain desired outcomes, as well as specific goals unique to each experiment. The first experiment models adversary decision-making using an Absorbing Markov chain (AMC). A sensitivity analysis of states (nodes in the graph) and actions (edges in the graph) is conducted which informs how downstream adversary decisions could be manipulated. The next experiment uses a Markov decision process (MDP). Assuming the adversary is initially blind to the rewards they will receive when they take an action, a Q´learning algorithm is used to determine the sequence of actions that maximizes the adversary rewards (called an optimum policy). This experiment gives insight in the possible decision-making of an adversary. Lastly, in the third experiment a two-player Markov game is developed, played by an agent (friend) and the adversary (foe). The agents goal is to decrease the overall rewards the adversary receives when it follows optimum policy. All experiments are demonstrated using specific examples.

Decision dominance—a Commander’s ability to see, think, understand, and act first—becomes exponentially more important in multi-domain operations (MDO) where the lethality and pace of warfare increase. At the heart of decision dominance is intelligence analysis. Intelligence analysis provides an understanding of the threat and environment that forms the foundation for Commanders’ decisions throughout planning and execution. Artificial intelligence and machine learning (AI/ML) offer opportunities to automate portions of the intelligence process. Accordingly, this paper explores mechanisms to employ machine learning (ML) techniques to rapidly synthesize heterogeneous text data into knowledge that can be graphically depicted in a situation template (SITEMP). Specifically, the paper examines the feasibility of two approaches. The first leverages hierarchical, agglomerative clustering with subsequent classification. This approach is analogous to how an analyst unfamiliar with the environment would operate—seeking patterns in the data to develop a fused understanding. The second approach applies a few-shot learning methodology that is akin to an analyst recognizing reporting based on prior experience. While AI/ML is not a panacea, it does promise significant gains to the competitor that can leverage it first and most fully. Ultimately, this paper seeks to inspire novel applications of AI/ML technologies to combat the challenges expected in MDO.

Effective human-robot teaming requires human and robot teammates to share a common understanding of the goals of their collaboration. Ideally, a complex task can be broken into smaller components to be performed by team members with defined roles, and the plan of action and assignment of roles can be changed on the fly to accommodate unanticipated situations. In this paper we describe research on adaptive human-robot teaming that uses a playbook approach to team behavior to bootstrap multi-agent collaboration. The goal is to leverage known good strategies for accomplishing tasks, such as from training and operating manuals, to enable humans and robots to “be on the same page” and work from a common knowledge base. Simultaneous and sequential actions are specified using hierarchical text-based plans and executed as behavior trees using finite state machines. We describe a real-time implementation that supports sharing of task status updates through distributed message passing. Tasks related to human-robot teaming for exploration and reconnaissance are explored with teams comprising humans wearing augmented reality headsets and quadruped robots. It is anticipated that shared task knowledge provided by multi-agent playbooks will enable humans and robots to track and predict teammate behavior and promote team transparency, accountability and trust.

In the DARPA Competency Aware Machine Learning program, the MindfuL™ system derived machine level ‘conditions’ (topics in a multi-modal Hierarchical Dirichlet Process model) that are aligned with and predictive of the competency of a CNN obstacle detection component of a self-driving car. These competency controlling conditions don’t map directly into human level concepts, which limits the utility of this explainable AI approach. This paper discusses methods to increase understanding and trust of the competency assessment and the ML agent itself by automatically generating labels for these conditions that will be meaningful and useful to the human operator.

Human and Autonomy Teaming (HAT) involves humans, artificial intelligence and machine learning (AI/ML), and autonomous systems (AS) working together to enhance performance and efficiency across a variety of fields. The purpose of this paper is to outline the current status, challenges, and perspectives related to HAT in the defense applications, such as multi-domain operations (MDOs). This paper emphasizes that, although ASs have numerous benefits, humans, and machines need to work together to ensure effective communication and coordination. Furthermore, this paper addresses the issues of situation awareness, HAT bias, security, privacy, confidentiality, trust, transparency, and explainability as they relate to joint HAT in MDOs. Finally, this paper explores the potential of perspectives to overcome these challenges and enhance collaboration between humans, AI/ML, and AS in the context of defense.

Knowing the future states of an adversary in an adversary-avoidance game can impart a survival advantage. To assess how predictive modeling helps agents achieve goals and avoid adversaries, we tested the efficacy of three predictive algorithms within a gridworld-based game. For one predictive algorithm, model predictions of adversary moves furnished to an agent helped the agent avoid capture compared to a case without predictions. A human-machine team scenario also benefited from model predictions, while humans alone experienced a ceiling effect. We investigated the efficacy of two additional predictive algorithms and present a maritime vessel pursuit scenario.

We explore strategies for improving the versatility of deep reinforcement learning (DRL) agents trained for maritime platform defense, in an effort to avoid impractical retraining when conditions change. DRL platform defense agents must be able to effectively schedule countermeasures, with constraints and against incoming raids of threats of uncertain type. Here we provide methods, centered on domain randomization, threat representation, and neural network architecture modification, for addressing changes in the relative locations and orientations of ships in the fleet, the number of ships and their inventories, and the distribution of threats. Testing our interventions in a realistic simulator, we show that a base DRL agent may be extended to account for a wide variety of changes in the operational scenario, with little degradation of performance.

This presentation was provided at SPIE Defense + Commercial Sensing 2023.

Conventional flight simulators usually include a complex and expensive outside vision projection system. Especially scenarios where the helicopter pilots look far to the side or through the windows near the pedals require a large projection dome to provide an image of the outside world. Additionally, simulators used for research need to be highly customizable: For rapid prototyping of new flight deck designs, the cockpit mockup must be adaptable enough to change the appearance and arrangement of its elements. The recent technological advancements of head-mounted displays (HMDs) offer many new ways to create a simulator that fulfills the stated requirements at moderate cost. A non-see-through HMD can immerse the pilots into a computer-generated cockpit with “unrestricted” virtual out-the-window view. The downwards view is even better than with dome projections. Such a fully virtual approach, however, requires complex finger-tracking and haptic feedback solutions to enable the user to interact with the cockpit. By contrast, a video-see-through HMD allows us to selectively combine a highly customizable virtual world with a video-stream of the real surroundings. One can, for instance, show the pilot’s hands and relevant parts of the physical flight deck mockup, enriched with virtual elements and virtual out-the-window view. In such a mixed setup, the pilots can naturally and directly interact with conventional input devices in an otherwise virtual environment. The paper presents our implementation of a mixed reality simulator with the Varjo XR-3 video-see-through HMD. We assess different variants, discuss implementation details like real-to-virtual-world-alignment, and explain the major challenges of such setups.

Within the U.S.-German project agreement on advanced technology for rotorcraft (PA–ATR) DLR conducts research on pilot aids for helicopter operations in DVE. For example, an HMD can provide visual cues to the pilot while flying eyes-out. With the HoloLens 2 as a low-cost HMD, the full-color waveguide display facilitates rapid prototyping of new color-coded symbology. To estimate the optical feasibility of the device, this paper presents findings on the see-through transmission of the combiner and a subjective rating of the color fidelity. The found issues can then be taken into account in the future design and development process of head-up display symbology.

This paper discusses how Augmented Reality (AR) and game engines can be utilized to increase situational awareness both within and beyond the battlefield. It presents a combination of software and hardware that allow for forward deployed operational forces and ground force commanders to automatically collect, share, and process data within their environment. The sharing of data produces a virtual representation of a real environment between multiple users, so that a common operating picture can be generated and processed. The use of the AR headset not only provides a way to present the data in a heads-up, eyes-out approach, but allows for interaction within the synthetic environment. Through this virtual interaction, the troops are able to annotate areas and people, improving non-verbal communication among a group. Areas can be marked as a destination and people can be marked as enemies. The annotations and virtualized environmental information can then be processed together to determine lines of sight of enemies and paths of least visibility to a destination point for all individuals wearing the AR devices. This processed data is displayed back to the troops so that they can make quick and informed decisions.

The power of artificial intelligence (AI) coupled with optimization algorithms can be linked to data-rich digital twin models to perform predictive analysis to make better informed decisions about installation operations and quality of life for the warfighters. In the current research, we developed AI connected lifecycle building information models through the creation of a data informed smart digital twin of one of US Army Corps of Engineers (USACE) buildings as our test case. Digital twin (DT) technology involves creating a virtual representation of a physical entity. Digital twin is created by digitalizing data collected through sensors, powered by machine learning (ML) algorithms, and are continuously learning systems. The exponential advance in digital technologies enables facility spaces to be fully and richly modeled in three dimensions and can be brought together in virtual space. Coupled with advancement in reinforcement learning and computer graphics enables AI agents to learn visual navigation and interaction with objects. We have used Habitat AI 2.0 to train an embodied agent in immersive 3D photorealistic environment. The embodied agent interacts with a 3D environment by receiving RGB, depth and semantically segmented views of the environment and taking navigational actions and interacts with the objects in the 3D space. Instead of training the robots in physical world we are training embodied agents in simulated 3D space. While humans are superior at critical thinking, creativity, and managing people, whereas robots are superior at coping with harsh environments and performing highly repetitive work. Training robots in controlled simulated world is faster and can increase their surveillance, reliability, efficiency, and survivability in physical space.

Commanders must remain agile and adaptive in the future Artificial Intelligence (AI)-enabled multi-domain battlespace, where critical decisions are made at the tactical edge. Over-reliance on static, cloud-centric approaches to Machine Learning Operations (MLOps) compromises such agility and adaptability. These systems must operate in a dynamic threat environment, and learn to detect novel threats during operation. They must be able to perform this learning through the execution of tactical MLOps under austere and degraded conditions, especially limited wireless network bandwidth. In response to these requirements, this paper describes Hawk, a system that leverages edge proximity for rapid and iterative execution of the Observe stage of the Observe-Orient-Decide-Act loop. Central to this architecture is the use of tactical cloudlets. These mini data centers provide cloud-like computing resources without the communication latency to exascale data centers. Hawk enables a human to guide MLOps at low cognitive load, thus enabling an operational objective to be achieved at speed and scale while remaining usable and explainable.

Artificial Intelligence applications are increasingly making use of vector embedding techniques to achieve impressive results in many application domains. Semantic Vector Spaces (SVS’s) are constructed using semantic vector embedding techniques that learn vector representations of data across multiple domains. An important application area enabled by such techniques is the capability to represent software services and service workflows as semantic hypervectors. Previous work has shown how these hypervector representations have significant advantages over alternative schemes for decentralized service workflow construction, particularly in low communications bandwidth and energy constrained environments that are typical in multi-domain operations. SVS construction usually assumes that all the data required to construct the semantic vector space is available centrally. However, in multi-domain operations different partners may not be willing to share the training data necessary to construct a common multi-domain SVS. Hence semantic hypervectors representing similar services or workflows but constructed from different training data by different partners cannot be discovered and used. In this paper we focus on how it is possible to map semantic hypervectors between partner SVS’s, so that complementary services and workflows developed and owned by different partners can be discovered and used to achieve mission goals. The paper describes techniques for generating the required mapping that require a minimum exchange of information between the different partners; demonstrate how it is possible to do this for semantic hypervectors that use different types of encoding (eg., real valued, binary, sparse slot-encoding); and illustrates how the mapping can be implemented in various multi-domain operational settings.

The Hierarchical Bayesian Neural Network (HNN) is a machine learning algorithm that attempts to use the natural hierarchical structure of data. HNN has demonstrated gains in robustness, accuracy, and reporting capabilities by addressing the technical challenge of classifying data at different levels of a hierarchical structure. There is a significant operational benefit in classifying at different levels of an ontology where the extracted knowledge is used for future decision-making, especially when classification at the finest level is unfeasible.

Traditional machine learning (ML) models used for enterprise network intrusion detection systems (NIDS) typically rely on vast amounts of centralized data with expertly engineered features. Previous work, however, has shown the feasibility of using deep learning (DL) to detect malicious activity on raw network traffic payloads rather than engineered features at the edge, which is necessary for tactical military environments. In the future Internet of Battlefield Things (IoBT), the military will find itself in multiple environments with disconnected networks spread across the battlefield. These resource-constrained, data-limited networks require distributed and collaborative ML/DL models for inference that are continually trained both locally, using data from each separate tactical edge network, and then globally in order to learn and detect malicious activity represented across the multiple networks in a collaborative fashion. Federated Learning (FL), a collaborative paradigm which updates and distributes a global model through local model weight aggregation, provides a solution to train ML/DL models in NIDS utilizing learning from multiple edge devices from the disparate networks without the sharing of raw data. We develop and experiment with a data-efficient, FL framework for IoBT settings for intrusion detection using only raw network traffic in restricted, resource-limited environments. Our results indicate that regardless of the DL model architecture used on edge devices, the Federated Averaging FL algorithm achieved over 93% accuracy in model performance in detecting malicious payloads after only five episodes of FL training.

Federated Learning (FL) enables collaborative model building among a large number of participants without revealing the sensitive data to the central server. However, because of its distributed nature, FL has limited control over the local data and corresponding training process. Therefore, it is susceptible to data poisoning attacks where malicious workers use malicious training data to train the model. Furthermore, attackers on the worker side can easily manipulate local data by swapping the labels of training instances to initiate data poisoning attacks. And local workers under such attacks carry incorrect information to the server, poison the global model, and cause misclassifications. So, detecting and preventing poisonous training samples from local training is crucial in federated training. To address it, we propose a federated learning framework, namely Confident Federated Learning to prevent data poisoning attacks on local workers. Here, we first validate the label quality of training samples by characterizing and identifying label errors in the training data and then exclude the detected mislabeled samples from the local training. To this aim, we experiment with our proposed approach on MNIST, Fashion-MNIST, and CIFAR-10 dataset and experimental results validated the robustness of the proposed framework against the data poisoning attacks by successfully detecting the mislabeled samples with above 85% accuracy.

Federated machine learning (FML) has proved a useful technique for training of artificial intelligence and machine learning (AI/ML) models, using data that is distributed among different constituents of a network which may be geographically dispersed. Typically, the data privacy of individual constituents should be preserved, and it may also be desirable to protect the integrity and secrecy of the algorithms and trained models deployed within the network. Demonstrating the privacy-enhancing technology of Confidential Computing, we present the results obtained using a novel solution for FML implementation that supports model training within a distributed network of data providers. Based upon recent research on the use of FML for distributed spectrum sensing in communication networks, we demonstrate the application of the proposed solution for distributed model training within a simulated sensor network of arbitrary topology. The presented solution provides for graph-based network configuration and model convergence within decentralized network applications. Cross-domain adaptation of the proposed solution and characteristics of confidential computing that support a zero-trust architecture are discussed, along with the integrated model integrity protection provided by attestation of trusted execution environments (TEEs). We conclude by looking ahead to the application of our solution to model training within distributed communications networks and sensor arrays, characterized by devices with limited electrical and computational power. We consider the use of physical unclonable functions (PUFs) to encrypt raw data before processing within a layered hierarchy secured with Confidential Computing technology.

Federated learning has been an active area of research for secure and privacy-aware learning where parameters and models are shared instead of actual data. However, there are some challenges to be addressed since federated learning relies on the aggregation process for the global model. If the learning server waits to receive all updates from all clients too long for the optimal global model, there might be a high delay. In this paper, we will design, develop and evaluate asynchronous federated learning by considering asynchronous client training and asynchronous aggregation that is applicable to tactical scenarios.

Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR) is a key technique in remote-sensing image recognition. The state-of-the-art Graph Neural Network (GNN) has achieved 99.09% accuracy on the MSTAR dataset. However, robustness is a critical concern when an adversary can add carefully crafted noise to the test image and mislead the classifier to give wrong output. In this work, we study the vulnerability of the state-of-the-art GNN-based SAR image classifier and show the feasibility of decreasing its accuracy by manipulating test images using the Fast Gradient Sign Method (FGSM). The intuition is that misclassifications are more likely to occur when the loss is high. For each SAR image, we add noise to every pixel, such that the perturbed image will be misclassified by the GNN model. For each pixel, the noise is determined by the gradient of the loss with respect to the pixel. Also, we guarantee that the noise values are sufficiently small such that the perturbation will be unnoticeable to human eyes. We conduct experiments by applying the attack on the 2425 test images from the MSTAR dataset. In our experiments, the accuracy of the target GNN drastically drops from 99.09% to 11.79% after we change each normalized pixel value (in [0, 1]) of the test images by less than 3/255.

Naval intelligence plays a critical role in multi-domain operations by identifying and tracking vessels of interest, especially suspected “dark ships” operating in an emissions-controlled (EMCON) state. While applying machine learning (ML) to maritime satellite imagery could enable an automated open-ocean search capability for dark ships, ensuring the robustness of ML models to environmental variations in the maritime domain remains a challenge because training sets do not encapsulate all possible environmental conditions. To address the challenge of unsupervised domain adaptation (UDA) in ship classification, i.e. transferring a ML model from a labeled source domain to an unlabeled target domain, we propose employing combinations of semi-supervised learning (SSL) techniques with standalone UDA approaches. Specifically, we incorporate combinations of FixMatch, minimum class confusion, gradient reversal, and mixup augmentation into the standard cross-entropy supervised loss function. These interventions were compared in two domain shift settings, one in which the source and target domains are both comprised of simulated data, and another in which the source domain consists of only simulated data, and the target domain consists of only real data. Experimental results comparing the combinations of interventions to a regularized fine-tuning baseline demonstrate that the greatest improvements in model robustness were achieved when combinations of our SSL strategy (FixMatch) and UDA algorithms were incorporated into training.

Reinforcement Learning continues to show promise in solving problems in new ways. Recent publications have demonstrated how utilizing a reinforcement learning approach can lead to a superior policy for optimization. While previous works have demonstrated the ability to train without gradients, most recent works has focused on the simpler regression problems. This work will show how a Multi-Agent Reinforcement Learning approach can be used to optimize models in training without the need for the gradient of the loss function, and how this approach can benefit defense applications.

As the Army begins to deliver on the promise of autonomous agents including unmanned ground and air vehicles and other intelligent agents the focus has shifted from functionality to robustness and resilience. Intelligent systems must not only be capable of performing specific tasks, but they must also do it in a manner that is resilient to the dynamism of real-world environments and adapt to the unexpected including adversaries. This paper surveys some of the recent research and breakthroughs towards robust and resilient autonomy, some of the keys to achieving systems that learn rapidly in new environments, and that demonstrate the ability deal with uncertainty through various means including leveraging human knowledge to greatly speed up the learning process.

The ability to accurately monitor the position of an object or group of objects in some video sequence can be of high value when making decisions with respect to security, maneuverability, ecology, and infrastructure. Efforts from our previous work^{1, 2} provided evidence that in overhead video, a system’s ability to perform this tracking could be greatly enhanced through effective and segmented calculation of frame-to-frame spatial correspondences. We now continue the investigations began by this previous work (wherein frames and their detections are co-mapped) through a case study wherein hand-crafted and machine learning (ML) approaches for co-registration of video frames are reviewed and compared. First, the merits and shortcomings of hand-crafted algorithms, ML models, and hybrid approaches are discussed for key-point detection and key-point description. Modifications to feature matching and homography estimation are discussed as well, and following this, a more recently published class of co-registration involving “detector-less” correspondence is outlined. These approaches are applied and evaluated on four overhead image sequences with ground-truth corresponding to object detection locations only. Because of this lack of ground control points for evaluation, co-mapped centroids of stationary objects are used to generate an accuracy metric for the various mapping approaches. Further, given the value of mapping and tracking in real-time contexts, this notion of accuracy is compared to computation time so that trade-offs in temporal and spatial performance can be better understood.

We present a framework for developing software agents via Machine Learning (ML) entitled Curriculum-Heavy Accelerated Learning in a Competitive Environment (CHALICE). CHALICE is designed to train and deploy intelligent agents capable of executing strategies for air-ground combat as embodied in AFRL’s MIST turn-based wargame system. Such agents can be used to suggest courses of action in real-time to operational planners and to provide an adversarial opponent for evaluation of proposed courses of action. CHALICE uses state-of-the-art Deep Neural Networks (DNNs) to represent the state of the environment and Deep Reinforcement Learning (DRL) to train each agent via repeated feedback from outcomes of the MIST Stratagem game. Unlike recent DRL approaches for strategy games such as Go or StarCraft [1] [2], CHALICE minimizes dependence on existing corpora of human gameplay and trains efficiently with low computational resources and short convergence time (hours to days rather than weeks to months). Over the course of four government-led competitions, CHALICE produced agents that continually improved their performance, resulting in competitive play against human and automated opposing agents at relatively low training cost and time. In this paper, we motivate the operational problem and technical challenges, provide an overview of our technical approach, elaborate on our vision-based and graph-based DNN architecture design and agent training procedure, and present results from the most recent Stratagem competition. We close with a discussion of future research recommendations.

Machine learning and artificial intelligence algorithms have expanded dramatically in use across diverse fields of research and practice. Despite the extensive benefits that these algorithms can bring to researchers, system designers, and operators alike, the adoption of these algorithms in space-related scenarios has lagged behind other fields. In order to encourage the increased adoption of artificial intelligence and machine learning techniques to space-domain-related problems, flexible modeling and simulation capabilities are needed to build stakeholder trust in these techniques. This research presents the development of a flexible Python-based modeling and simulation environment for applying Reinforcement Learning to Low Earth Orbit satellite Hyper Spectral Imaging sensor tasking. With the transition away from small numbers of highly exquisite on-orbit systems to proliferated architectures characterized by constellations of lower cost and complexity spacecraft, the methods by which payload sensors are tasked have become dynamic and complex, making the problem of determining effective sensor tasking methods an important area of research. Such a problem lends itself well to the application of Reinforcement Learning. The focus of this work is on developing the role of intelligent systems in improving the data acquisition process in a space-based hyperspectral imaging system, and showing how the developed modeling and simulation framework can be successfully employed to improve the acquisition of targets of interest. A key strength of the presented reinforcement learning application framework is its non-commercial, extensible nature, suitable for both research and educational purposes.

Recently, deep reinforcement learning (DRL) algorithms have been adapted for real-time control and policy-based decision for robots, drones, and autonomous vehicles. Traditional implementations of DRL use general purpose computing (CPU) and graphical processing units (GPU). In current work, we present HardCompress as an optimized hardware configuration for neural network (DNN) accelerators using High Level Synthesis (HLS) techniques.

There is a growing orientation of cyber systems, technologies, and processes away from notions of cybersecurity and towards notions of cyber resilience. In multi-domain operation settings, such as joint all-domain command and control (JADC2), the scale and evolvability of networks makes resilience an especially useful concept for cyber operations, as the attack surface is simply too complex to secure. Generic design patterns for cyber resilience in cyber-physical systems have been difficult to identify. This is due in part to the generality of cyber-physical systems (CPS). Artificial intelligence and machine learning (AI/ML), however, are more specific than CPS, typically serving as a component or subsystem within a CPS, and thereby offer a more promising opportunity to define general resilience concepts for AI/ML-based cyber systems. In this paper, we introduce resilience concepts for AI/ML-based cyber systems. We consider the measurability and testability of resilience, as well as suggest possible requirements for resilience. To illustrate and contextualize the discussion of resilience, we use a notional example of AI/ML-based network defense analysis in a JADC2 setting. With the aid of the JADC2 example, we discuss the difficulties in engineering AI/ML for cyber resilience. Although AI/ML often have learning processes embedded, that does not make them resilient, as the learning processes are dependent on other systems, technologies, and processes. This paper concludes that to extend cyber resilience to an AI/ML-based cyber systems, one can use generic design patterns to identify resilience mechanisms and to associate them with testable, measurable requirements.

Planning the delivery of cargo as part of an airlift operation is a notoriously complex problem. Transportation routes can suddenly become inaccessible due to poor weather or other unexpected occurrences. Factors such as airplane speed, carrying capacity, and airport maximum-on-ground must also be considered to ensure on-time and efficient delivery. Unforeseen cargo needs may require quick re-planning to meet tight deadlines. To address the problem, we held the Airlift Challenge, an online multi-agent planning competition which concluded in early 2023. Competitors in the Airlift Challenge created innovative algorithms to execute a simulated airlift operation. The algorithms were scored against a set of increasingly complex evaluation scenarios while contending with unexpected events and disruptions. In this paper, we describe the competition and simulation environment, summarize results, and include write-ups of the top approaches provided by the winning teams.

Despite the progress in deep learning networks, efficient learning at the edge (enabling adaptable, low-complexity machine learning solutions) remains a critical need for defense and commercial applications. We have pursued multiple neuroscience-inspired AI efforts which may overcome these data inefficiencies, power inefficiencies, and lack of generalization. We envision a pipeline to utilize large neuroimaging datasets, including maps of the brain which capture neuron and synapse connectivity, to improve machine learning approaches. We have pursued different approaches within this pipeline structure, including data-driven discovery in biological networks, augmenting existing computational neuroscience models, investigating the biological structure which enables behaviors, and modifying existing machine learning architectures with insight from biological structure. First, as a demonstration of data-driven discovery, the team has developed a technique for discovery of repeated subcircuits, or motifs. These were incorporated into a neural architecture search approach to evolve network architectures. Second, we have conducted analysis of the heading direction circuit in the fruit fly, which performs fusion of visual and angular velocity features, to explore augmenting existing computational models with new insight. Our team discovered a novel pattern of connectivity, implemented a new model, and demonstrated sensor fusion on a robotic platform. Third, the team analyzed circuitry for memory formation in the fruit fly connectome, enabling the design of a novel generative replay approach. This replay approach resulted in an over 20% accuracy improvement in an incremental class learning scenario, and also demonstrated the ability to utilize large neuroscience datasets to analyze the neural connectivity underlying behavior. Finally, the team has begun analysis of connectivity in mammalian cortex to explore potential improvements to transformer networks. These constraints increased network robustness on the most challenging examples in the CIFAR-10-C computer vision robustness benchmark task, while reducing learnable attention parameters by over an order of magnitude. Taken together, these results demonstrate multiple potential approaches to utilize insight from neural systems for developing robust and efficient machine learning techniques.

K-Means is one of the most popular clustering techniques and has been successfully used in many data mining fields. K-Means is computationally expensive for high dimension and high quantity data and various techniques have been developed to reduce the computational cost. These techniques mainly involve improving implementation efficiency. In this paper, we proposed a Hausdorff distance and diffusion equation evolution combined technique to speed up K-Means through reducing number of data points in distance computation. Experiments show that for manifold data, the proposed method not only significantly reduces the computational cost, but also improve the clustering performance.

An ongoing challenge for image processing algorithms is how to go beyond identification, detection, and classification. This idea encompasses the ability to reason about with contextual information. Including what the whole image scene represents and how it informs. The problem of current algorithms may include approaches that only focus on a single focal plane discarding peripheral elements, limitations of datasets where salient objects are captured close up but not from different angles or distances. Our research explores how to leverage not just core features but features that may have been assumed to be spurious in nature. Thus contextual information increase performance of algorithms: accuracy, interpretability and meaning. In this paper we will present research into causal reasoning and stable diffusion cite{rombach2021highresolution} utilizing synthetic data for detection with core and spurious features.

In many defense settings, agents are on the same team as the central decision-maker; they express preferences truthfully to that decision-maker, who then aggregates those expressed preferences to decide on an outcome. This strays from, and is easier to analyze than, the traditional mechanism design setting from microeconomics, where agents are not assumed to be truthful. But, what if those trusted agents’ true preferences are manipulated by upstream actors (either adversaries or nature)? How should a decision-maker act when either strategic or natural uncertainty manipulates trusted agents’ beliefs?

With numerous technologies, seeking to utilize deep learning-based object detection algorithms, there is an increased need for an innovative approach to compare one model to another. Often, models are compared one of two over-arching ways: performance metrics or through statistical measures on the dataset. One common approach for training an object detector for a new problem is to transfer learn a model, often initially trained extensively on the ImageNet dataset; however, why one feature backbone was selected over another is overlooked at times. Additionally, while whether it was trained on ImageNet, COCO, or some other benchmark dataset is noted, it is not necessarily considered by many practitioners outside the deep learning research community seeking to implement a state-of-the-art detector for their specific problem. This article proposes new strategies for comparing deep learning models that are associated with the same task, e.g., object detection.

Internet of Things (IoT) devices have communication, computation, size, weight, battery life, and power consumption limitations. Machine learning (ML) algorithms on IoT devices suffer from limited computational resources, resulting in poor performance. Techniques such as model pruning, offloading, and data compression can improve computational, network, and storage costs as well as performance at the expense of inference quality. This study examines the performance of ML inferencing tasks and adaptations on IoT devices using statistical learning methods such as ridge regression and random forest. We aim to understand the trade-offs of inference adaptations in a range of operational regimes including constraints on available bandwidth, distance from the data, and constraints on compute. Our results indicate that the task configuration, offload or on-device, and the bandwidth available are the most critical factors in determining inference performance, while the percentage of model pruned is the least important. These findings demonstrate how statistical learning can be used to better understand the effects of task performance on IoT device and offers insights into which inference adaptations provide the largest improvements in ML inference tasks to support real-time requirements.

Ensuring the realism of computer-generated synthetic images is crucial to deep neural network (DNN) training. Due to different semantic distributions between synthetic and real-world captured datasets, there exists semantic mismatch between synthetic and refined images, which in turn results in the semantic distortion. Recently, contrastive learning (CL) has been successfully used to pull correlated patches together and push uncorrelated ones apart. In this work, we exploit semantic and structural consistency between synthetic and refined images and adopt CL to reduce the semantic distortion. Besides, we incorporate hard negative mining to improve the performance furthermore. We compare the performance of our method with several other benchmarking methods using qualitative and quantitative measures and show that our method offers the state-of-the-art performance.

The output of a semantic segmentation model on an off-road dataset can provide an accurate description of the terrain and the obstacles contained within it. This output can be leveraged to determine the presence of barriers in an image. An obstacle is anything that may obstruct a portion of the region of traversal, while we define a barrier as something that will bisect the region of traversal to create two disjoint regions that would otherwise be connected if not for its presence. Detecting instances of barriers requires more than learning the correct label for a standard 2D semantic segmentation model. This paper will present an approach to detect the presence of barriers/barricades in the scene by utilizing the traversability of the semantic classes of non-traversal and the pose of that class(es) in relation to other classes of non-traversal in the scene to define an object as a barricade/barrier.

In this article, we present SFMLOps, a Security Framework for Machine Learning Operations (MLOps), a comprehensive and novel approach to securing MLOps pipelines in multi-domain operations. SFMLOps can be used to benchmark security in mobile cyber-physical systems like quadruped reconnaissance robots, unmanned autonomous vehicles, and wearable brain-computer interfaces. Our framework examines and categorizes potential attack surfaces and threats within MLOps, offering countermeasures and their effectiveness for various aspects such as data storage, communication channels, model training, and model predictions. We provide security engineers practical guidance on developing secure MLOps pipelines. We introduced a secure pipeline design, MLPipeSec, based on a publisher-subscriber model and implemented on the JointForceNet+ Blockchain to ensure end-to-end trustworthiness across the MLOps pipeline. To evaluate the impact of security parameters on a multi-domain computer vision task, we compared several frameworks for their security and performance using the CIFAR-10 dataset. We also investigated Gossip Learning as a federated learning framework in conjunction with Google Cloud Platform and introduced a new federated learning model, VizFedML. Our experimental results demonstrate the efficacy of the SFMLOps framework and the MLPipeSec design in mitigating a range of vulnerabilities and weaknesses associated with MLOps, contributing to the development of more secure and robust machine learning systems.

In recent years, computer vision has made significant strides in enabling machines to perform a wide range of tasks, from image classification and segmentation to image generation and video analysis. It is a rapidly evolving field that aims to enable machines to interpret and understand visual information from the environment. One key task in computer vision is image classification, where algorithms identify and categorize objects in images based on their visual features. Image classification has a wide range of applications, from image search and recommendation systems to autonomous driving and medical diagnosis. However, recent research has highlighted the presence of bias in image classification algorithms, particularly with respect to human-sensitive attributes such as gender, race, and ethnicity. Some examples are computer programmers being predicted better in the context of men in images compared to women, and the accuracy of the algorithm being better on greyscale images compared to colored images. This discrepancy in identifying objects is developed through correlation the algorithm learns from the objects in context known as contextual bias. This bias can result in inaccurate decisions, with potential consequences in areas such as hiring, healthcare, and security. In this paper, we conduct an empirical study to investigate bias in the image classification domain based on sensitive attribute gender using deep convolutional neural networks (CNN) through transfer learning and minimize bias within the image context using data augmentation to improve overall model performance. In addition, cross-data generalization experiments are conducted to evaluate model robustness across popular open-source image datasets.

Artificial intelligence (AI) and machine learning (ML) systems are required to be fair and trustworthy. They must be capable of bias detection and mitigation to achieve robustness. To this end, a plethora of research fields have seen growth in research related to making AI/ML systems more trustworthy. Causal learning and Explainable AI (XAI) are two such fields that have been used extensively in the past few years to achieve explainability and fairness. However, they have been used as separate methodologies, not together. This paper provides a new perspective in using causal learning and XAI together to create a more robust and trustworthy system. Having causality and explainability together in the same model presents an extra layer of robustness, that is not achieved by using either of them individually. We present a use case for combining causality via causal discovery, and explainability via feature relevance. Using causal discovery, the generated causal graphs are compared to the feature relevance plots from the ML model. Directed causal graphs can display the features that are causally relevant for the predictions, and these causally relevant features can be directly compared to the features listed from correlation-based explanations from XAI.

Multi-Agent Reinforcement Learning (MARL) is a key technology in Artificial Intelligence Systems. The development of MARL is a “Training-in-Simulation” process and involves time-consuming computation and communication. Existing implementations of MARL using homogeneous architecture fail to satisfy the above requirements simultaneously. In this work, using a state-of-the-art MARL algorithm - Multi-Agent Deep Deterministic Policy Gradient (MADDPG) - as a benchmark, we provide a systematic analysis of the performance bottleneck on multi-core CPU platforms. We show that (a). the training throughput of large neural network models can benefit from data parallel architecture such as GPU; (b) Memory-intensive components including replay management and all-to-all communication leads to large memory overheads on multi-level cache memory hierarchy on CPU and GPU, thus should be accelerated using custom hardware coupled with large on-chip SRAM with specialized data layout optimizations on FPGA.

Demand for 3D X-ray CT inspection (in the medical, nondestructive testing, and security fields) is increasing year by year, and AR (Augmented Reality), VR (Virtual Reality), and MR (Mixed Reality) are being used for displaying internal structure of object. Although, these applications can display surface-rendered or volume-rendered objects, but have not been developed to accurately represent and spatially help the observer understand the point he or she wants to see. In this research, the representation method and system were proposed that 2D image output from DICOM is superimposed on the cross-section of a surface-rendered model, and by using a motion sensor, an object imaged with a 3D X-ray CT can be freely manipulated in the virtual space using the human hand, just as if the object were actually moved in real space, so the cross-section of the point to be checked can be viewed in any direction. By using the stencil buffer function, a shader function on Unity, unnecessary areas other than the cross-section can be hidden, so the object itself is not hidden in the 2D image. This system can be performed with either spatial reality display, VR device, or AR glasses. The results show that the proposed method is effective as a useful 3D representation in 3D X-ray CT, which is voxel data containing internal information, and as a method that allows viewing tomographic images in CT images and easily representing specific locations.

In this paper, we introduce HDA, an iterative hyperplane-search method for Generalized Discriminant Analysis (GDA), designed to tackle challenges associated with high-dimensional data and limited sample sizes. Traditional LDA and GDA methods often struggle with poor covariance estimates, which lead to overfitting and reduced generalization capabilities. Furthermore, standard single-shot GDA incurs a high computational cost, particularly for high-dimensional data. Our proposed HDA method provides a tunable performance-complexity trade-off, enabling better model generalization by iteratively refining the discriminant analysis solution. We extensively evaluate HDA’s performance on various benchmark datasets and compare it to other GDA and LDA alternatives. The experimental results highlight the merits of HDA in terms of classification performance, balance between computational complexity, and enhanced model generalization. This study paves the way for improved discriminant analysis techniques, particularly in scenarios where data dimensionality and limited sample sizes pose significant challenges.

Adversarial training (AT) is considered the most effective strategy to defend a machine learning model against adversarial attacks. There are many different methods to perform AT, but the underlying principle is the same, namely, augment the training data with adversarial examples. In this work, we investigate the efficacy of four different adversarial example generation strategies on AT of a given classification model. The four methods represent different categories of attack and data. Specifically, two of the adversarial generation algorithms perform attacks in the pixel domain, while others operate in the latent space of the data. On the other hand, two of the methods generate adversarial data samples designed to be near the model decision boundaries, while the other two generate generic adversarial examples (not necessarily at the boundary). The adversarial examples from these methods are used to adversarially train models on MNIST and CIFAR10. In the absence of a good metric to measure robustness of a model, capturing the effect of AT using a single metric can be a challenge. Hence, we opt to evaluate the robustness improvements resulting of the adversarially trained model using a variety of empirical metrics introduced in the literature that measure local Lipshitz value of a network (CLEVER), smoothness of decision boundaries, robustness to adversarial perturbations and defense transferability

Cyber Physical Systems (CPS) security within industrial fields have enforced itself, due to its deployment critical infrastructure. The complexity, and diversity are evolved with these CPS systems. While connectivity demands for these systems to communicate with each other increases, their attack surface expands. The impact of cybersecurity has on business continuity increase. ICS can run real time critical function, where firewall inspection delay can fail the process. Hence a special firewall consideration needs to be implemented. The Uptime requirements for these ICS is extremely high, which means the normal maintenance or security patches is out of discussion. Beside a modification on any ICS due to firmware update, can trigger revalidations for all interconnected ICS. Perimeter defense firewall is one of the common strategies to protect these ICS systems. The firewall will inspect and detect ingress traffic. Internal firewall will be more enhanced way to protect also from internal attacks within the network. Hence, a need for more efficient ways to detect these attacks, based on Deep Learning (DL) approach with a good source of (Industrial Internet of Things) IIoT dataset. This conference paper evaluates Deep learning approach using Bi-directional Long Short-Term Memory (BLSTM) on resent publicly available dataset “Edge-IIoTset”. This dataset has realistic dataset of IIoT applications. With more than 10 types of sensors/devices uses in ICS systems. With fourteen attacks including DoS/DDoS attack. In this research paper, we consider utilizing deep learning algorithms (BLSTM) to detect and protect the service availability of Critical Infrastructure (CI) and Industrial Control Systems (ICS) from Denial of Service (DoS)attack. The research proposal considers most recent dataset with packet compared with flow format to train our module. The benchmarking with common metrics is used as baseline to compare algorithm efficiency, where accuracy of 99.877 was achieved and validation time of 18millisconds.

Short Messaging Service (SMS) becomes a more easy, affordable way to communicate and increasingly replace phone calls. Spam is any kind of unwanted, random unsolicited message that gets sent without any authorization from the receiver. hackers use spam SMS to get their important information. Effective spam detection is an essential tool for assisting users in determining whether an SMS is a spam or not. Different machine learning methods such as Deep Learning techniques have attempted to distinguish between spam and ham SMS texts. This paper proposes the Spam-Ham Classification method using Recurrent Neural Networks (RNN) and Long Short-Term Memories (LSTM). The proposed model utilizes Keras and TensorFlow to detect Spam SMS. The dataset used is SpamSMSCollection from the UCI machine learning repository. The dataset contains a set of 5574 SMS messages. The dataset is preprocessed using tokenization, Lemmatization, padding, and stopword removal. The overall accuracy of the proposed model is 98%. The performance of the proposed method is compared with different machine learning algorithms such as Support Vector Machine (SVM), K-nearest neighbors (KNN), and Multi-layer Perceptron (MLP).