4
Network, Cyber, and Computational Sciences

INTRODUCTION

The U.S. Army Combat Capabilities Development Command (DEVCOM) Army Research Laboratory’s (ARL’s) network, cyber, and computational sciences competency focuses on distributed, resilient, secure networking and resource-adaptive decentralized computing for decision dominance. The work within the competency focuses on the joint study of networking, cybersecurity, and computational sciences. Key features of the competency include the following: multi-disciplinary and cross-competency efforts to enable and accelerate future capabilities; intelligent, data-centric networking to maximize analytics, artificial intelligence (AI) capacity and resilience; joint adaptation of network, computation, and analytics “to the edge” and not “at the edge;” employment of universal computational methods to accelerate technology advancements, developing of responses to adversaries; and leveraging quantum networking to avoid technological surprise.¹

Within the competency are three core competencies—computational methods for modeling and learning, cyber defense and cybersecurity, and resilient and adaptive communication networks. The computational methods for modeling and learning core competency focuses on the development of mathematical algorithms, deterministic and stochastic models, multi-scale methods, and uncertainty quantification to simulate complex, physical systems to understand variability, predicting system evolution with quantified confidence, and enabling exploitation of those predictions in both high-performance and limited-compute settings. It also focuses on developing and leveraging deep learning techniques to tackle an ever-expanding array of sensing applications involving acoustic, electromagnetic, radio frequency, and optical modalities. The cyber defense and cybersecurity core competency focuses on theories, models, optimized algorithms, and experimentation for preventing, detecting, mitigating, monitoring, and predicting of adversarial activities and their impacts within cyberspace. It also focuses on the development of cyber deception and counter-deception strategies to provide resilience and defenses against adaptive and sophisticated adversaries. Its scope includes traditional enterprise-level and tactical information networks and non-traditional networks such as communication buses found on vehicle platforms. The resilient and adaptive communication networks core competency focuses on theories, methods, algorithms, and experimental approaches to enable resilient communications in complex and contested environments via novel communication modalities, multi-layer adaptive protocols for robust information delivery (including storage, computing, and communications), emerging quantum networks, interpretable and adversarial machine learning (ML) to enable autonomous control of heterogeneous network structures and dynamics for resilience to adversarial attacks.

On August 6–8, 2024, the Panel on Assessment of Network, Cyber, and Computational Science visited the Adelphi Laboratory Center in Adelphi, Maryland. During this visit, the panel viewed podium and poster presentations, toured facilities, and spoke to the scientific researchers within the competency. Below is the summary of the panel’s findings as they relate to the assessment criteria. As part of the

___________________

¹ Passages of these paragraphs from B. Rivera, 2024, “Network, Cyber & Computational Sciences Competency Overview,” U.S. Army Combat Capabilities Development Command (DEVCOM) Army Research Laboratory (ARL) presentation to the committee, August 6.

assessment, the Army Research Laboratory Technical Assessment Board (ARLTAB) and its panels were asked by ARL to provide suggestions of specific people and organizations relevant to the work it is doing, or could be doing, that ARL could connect with. The fruits of this brainstorming activity are captured in the pages of the following chapter. It is important to note, however, that while certain individuals or organizations are mentioned, the ARLTAB and its panels are in no way being prescriptive about connecting to these outside entities and understand there may be other exemplars of equal value in the research community. ARL should therefore use its best judgement as to whether these ideas could be helpful.

TECHNICAL QUALITY OF THE CORE COMPETENCIES

Computational Methods for Modeling and Learning Core Competency

Achievements and Advancements

The research within the computational methods for modeling and learning core competency was organized into two areas: (1) computational modeling of complex systems and (2) universal computational methods. Overall, taken individually, the strongly empirical and applied focus of intramural ARL projects was appropriate and reasonable. While the research may initially appear incremental concerning work in the broader research community, there is a dearth of strongly applied research offering practical insights on what works and what does not in typical, realistic scenarios of interest. Such questions are not typically pursued by academia due to the perceived lack of novelty. Even industry laboratories do not always address the same kinds of application scenarios being considered at ARL. Thus, ARL research fulfills an important and potentially impactful role.

ML is a cross-cutting theme featured prominently in all three core competencies. Several projects in other core competencies employed the use of toy data sets and experiments that were comparatively more divorced from reality. By contrast, many of the research projects within this core competency employed highly realistic experimental setups and evaluations, with some projects even progressing toward field experiments with systems like Husky UGV (unmanned ground vehicle).² The practical usefulness and potential impact of this work across other ARL projects (e.g., Resource Adaptive Multidomain ISR to the Tactical Edge [RAMITE]) were impressive. ARL’s development and use of testbeds offer an excellent mechanism for evaluation and demonstration of research outcomes. The approach has the additional benefit of informing the researchers to formulate problems based on realistic assumptions. Stated differently, the flow of ideas and information between algorithm researchers and system developers is bi-directional.

More specifically, concerning the various presentations on “accelerated inference” and “adaptive inference” (project names: Adaptive Inference Framework of Video Object Detection on Edge Devices for Dynamic Environments and Methods for Accelerated Inference of Neural Networks on Resource Constrained Devices of Neural Networks on Edge Devices), the projects utilized empirical approaches to evaluate the research outcomes. This is appropriate for the projects for two reasons: first, characterization of an execution platform requires a certain amount of empirical evaluation; and second, the broader ML field does not currently have a well-developed underpinning theory to support an analytical approach to the problem. However, the utilized empirical methodology could be strengthened by looking beyond the obtained experimental outcomes at a high level; for example, the runtime of a given model variant on a specific execution platform, to develop an understanding of its constituting elements. Such an understanding is expected to enable the development of simple estimators for the prediction of execution

___________________

² Clear Path Robotics, “Husky A300 Unmanned Ground Vehicle,” https://clearpathrobotics.com/husky-unmanned-ground-vehicle-robot, accessed October 16, 2024.

times of a new ML model workload on a new execution platform. Examples of such work exist in the embedded systems community.

The intramural presentation “Continuous Time Digital Signal Processing Disruptive Advantages” was extremely interesting and represented innovative and potentially high-impact work. This work represents a high-risk, high-reward direction at the forefront of technology for low-power sensing and edge computing. The concept is scientifically innovative, although investigation of its trade-off in the context of target workload is not sufficiently advanced yet. The approach could enable significant power savings at the edge, thereby, bringing about considerable improvement in size, weight, and power (SWaP) metrics. Admittedly, the tools and hardware for asynchronous processing are not well-developed; however, a rough approximation of the potential energy and memory savings should be doable with modest effort, to de-risk the concept and evaluate if further investigation is justified. In the longer term, event-driven sensing would naturally tie into and complement the work on spike-based ML models, which is being investigated in other projects. Moreover, there is considerable opportunity to advance this line of research. In particular, a careful mathematical analysis, from the perspective of approximation theory and even information-based complexity, could elucidate the classes of signals (functions) the proposed method will achieve compression and by how much; in other words, one might be able to construct useful bounds or guarantees.

The work and methodology described in “Adaptive LPD Radar Waveform Design with Generative Deep Learning Resource Constrained Devices” was well-motivated and well-constructed, and the presentation of this work was outstanding. Although there are existing applications of generative adversarial networks (GANs) to waveform synthesis in addition to previous work on physics-informed loss with GANs,³ the application of GANs to generate low probability of detection (LPD) radar waveforms without labeled data represents a unique, interesting, and potentially impactful research contribution.

Finally, the field of ML has a strong tradition of open science. During the review, it was encouraging. The research on multiscale modeling was one commendable example of this. Moreover, authors of other research projects such as Methods for Accelerated Inference of Neural Networks on Resource Constrained Devices and Generalizing Out-of-Distribution Using Deep Reinforcement Learning expressed an interest in open-sourcing their work in the future. While open-sourcing software benefits the larger research community and facilitates reproducibility of the work, it can also help to improve published ideas and spark future research directions. ARL’s guidance on open-sourcing software, published to ARL’s GitHub repository,⁴ was also impressive and serves as a shining example of how a government laboratory can responsibly open-source software under necessary constraints related to security.

Research Portfolio Opportunities

There is a broad range in the quality of the presented research, and this section seeks to provide some suggestion on where improvements can be made. While some of the scientific results of the projects appear to be promising, the level of rigor in the presented work for the projects Continuous Time Digital Signal Processing Disruptive Advantages and Parametric Uncertainty Quantification: From Stochastic Galerkin to Deep Generative Learning, and Multiscale Modeling of Architected Materials could be improved through more analysis to corroborate the simulations. This could help bolster confidence in the results and potentially facilitate adoption by the broader research community.

___________________

³ See, for example, L. Yang, S. Treichler, T. Kurth, K. Fischer, D. Barajas-Solano, J. Romero, V. Churavy, A. Tartakovsky, M. Houston, and G. Karniadakis, 2019, “Highly-Scalable, Physics-informed GANs for Learning Solutions of Stochastic PDEs,” arxiv, https://doi.org/10.48550/arXiv.1910.13444.

⁴ ARL, “ARL Open-Source Guidance and Instructions,” https://github.com.mcasgov.us/USArmyResearchLab/ARL-Open-Source-Guidance-and-Instructions, accessed October 16, 2024.

There was also a slight risk across projects of being duplicative of existing work. Enhancing efforts to review open-source research may help avoid duplicative or similar research efforts. The following projects would benefit from better delineation from previous research:

• In the Neural Network Modifications for Training at the Edge project, aside from being comparatively less of a research-focused project than others, the work focused exclusively on training a classifier for radio waveforms, which is not novel. There are a number of papers^5,6 and blog posts^7,8 on the subject.
• In the project Deep Neural Network Sensitivity to Varied Radar Waveform Properties for Radar Classification and Domain Adaptation, the team presented a synthetic radar waveform generator called pyRadarWF. While the work is quite interesting and important with initial promising results, it was not adequately distinguished from existing work in this space. For instance, the problem being addressed and the proposed solution appear similar, if not identical, to Wu et al. (2023).⁹ Moreover, there are a number of both commercial and open-source synthetic data generators in the radar domain in addition to proposed solutions in academia. Examples include Ansys,¹⁰ radarsimpy,¹¹ and Simluated Radar Waveform Generator by the National Institute of Standards and Technology (NIST).¹² The project would benefit from establishing a clearer distinction between pyRadarWF and existing synthetic radar generators.
• The work on using generative diffusion modeling for uncertainty quantification (UQ) (project name: Parametric Uncertainty Quantification: From Stochastic Galerkin to Deep Generative Learning) is interesting. However, it is not the only recent approach of using operator learning for performing UQ.^13,14
• The work on using diffusion maps to encode physical laws through geometry (project name: Geometry-Driven Surrogate Modeling of RDX Chemical Kinetics) is also interesting, although similar ideas have been previously described.¹⁵
• In “Generalizing Out-of-Distribution with Deep Reinforcement Learning,” a benchmark environment to evaluate reinforcement learning (RL) models in the face of out-of-distribution (OOD) examples is presented. The benchmark environment augments the existing PROCGEN environment to generate OOD examples, and the work and experiments in this

___________________

⁵ M. Zhang, M. Diao, and L. Guo, 2022, “Convolutional Neural Networks for Automatic Cognitive Radio Waveform Recognition,” IEEE Access 5:11074–11082.

⁶ T. Wang, G. Yang, P. Chen, Z. Xu, M. Jiang, and Q. Ye, 2022, “A Survey of Applications of Deep Learning in Radio Signal Modulation Recognition,” Applied Sciences 12:12052.

⁷ LinkedIn, 2020, “Classifying RF Signals Using Machine Learning,” updated September 20, https://www.linkedin.com/pulse/classifying-rf-signals-using-machine-learning-petros-mouchtaris.

⁸ Panoradio SDR Advanced Software Defined Radio, 2023, “Practical RF Machine Learning for Signal Recognition,” updated March 23, https://panoradio-sdr.de/practical-rf-machine-learning-for-signal-recognition.

⁹ H. Wu, X. Li, L. Lan, L. Xu, and Y. Tang, 2023, “Towards Radar Emitter Recognition in Changing Environments with Domain Generalization,” avXiv, https://doi.org/10.48550/arXiv.2302.09359.

¹⁰ Ansys Blog, 2023, “Simulating Reality: The Importance of Synthetic Data in AI/ML Systems for Radar Applications,” updated February 8, https://www.ansys.com/blog/synthetic-data-in-ai-ml-radar-systems.

¹¹ Radarsimx, “Radarsimpy,” https://github.com/radarsimx/radarsimpy, accessed October 17, 2024.

¹² National Institute of Standards and Technology, “Simulated Radar Waveform Generator,” https://github.com/usnistgov/SimulatedRadarWaveformGenerator, accessed October 17, 2024.

¹³ L.F. Guilhoto and P. Perdikaris, 2024, “Composite Bayesian Optimization in Function Spaces Using NEON—Neural Epistemic Operator Networks,” arxiv, https://doi.org/10.48550/arXiv.2404.03099.

¹⁴ C. Moya, A. Mollaali, Z. Zhang, L. Lu, and G. Lin, 2025, “Conformalized-DeepONet: A Distribution-free Framework for Uncertainty Quantification in Deep Operator Networks.” Physica D: Nonlinear Phenomena 471:134418, https://doi.org/10.1016/j.physd.2024.134418.

¹⁵ B.E. Sonday, M. Haataja, and I.G. Kevrekidis, 2009, “Coarse-Graining the Dynamics of a Driven Interface in the Presence of Mobile Impurities: Effective Description via Diffusion Maps,” Physical Review E 80:031102.

• project were commendable. It should be noted that Mohamed and Valdenegro-Toro (2021)¹⁶ also developed a benchmark environment for OOD detection with corrupted observed states or modified physical parameters. It was not completely clear what additional insights on RL models the proposed benchmark environment described by ARL was able to provide. Work by Nasvytis and colleagues¹⁷ is another more recent example of an OOD benchmark environment (see reference for code).¹⁸ While these works focus on OOD detection, perhaps an experimental comparison to existing OOD benchmark environments for RL might be beneficial prior to publication of this work.

Additionally, while individual research areas within the competency have been met with success, as described in the “Accomplishments and Advancements” section above, a challenge to assessing the overall research quality of the core competency is that the portfolio of scientific projects appears to be more of a collection of results sown across different areas of computational methods without a coherent scientifically focused framework. As a result, the scientific topics did not seem to connect or cross-fertilize in meaningful ways. In addition, half of the presented portfolio (e.g., much of the work under “universal computational methods”) focused on engineering adaptations of neural networks or on the development of benchmarking systems. In flavor, this felt more like an extension of the RAMITE project than the development of computational methods. Thus, ARL has the opportunity to develop a scientific framework that incorporates broader themes and challenges in contemporary mathematics as they relate to the scientific goals of the core competency. This will help ARL build a more rigorous state-of-the-art portfolio that could allow greater connections and cross-fertilization between projects.

To develop this framework, both the intramural and extramural PSQs could be grouped or reshaped into fewer, more persistent and foundational themes and topic areas that stay in alignment with the core competency’s scientific goals and the goals defined for both the intramural and extramural programs and describe the scientific challenges and core questions ARL is seeking to address. This framework may also consider the relevant scientific challenges within other core competencies if doing so creates more integration and cross-core-competency support.

Finally, ARL may want to re-evaluate its plans for de-emphasizing statistics and optimization given that the foundational work in ML, perhaps the largest driver of technical innovations at the moment, is rooted in statistics and optimization.

Future Research Directions

Several intramural research projects focused on topics such as adaptive video object detection, accelerated neural network inference on edge devices, and the handling of out-of-distribution data. These topics are currently being treated as distinct and separate. However, from another perspective, they can be viewed as interconnected through the area of federated AI techniques, such as federated learning and federated agents. For instance, re-training or fine-tuning a model via federated learning on newly collected data has been investigated as a proposed solution to handling distribution shifts in resource-

___________________

¹⁶ A.P. Mohammed and M. Valdenegro-Toro, 2021, “Benchmark for Out-of-Distribution Detection in Deep Reinforcement Learning,” arxiv, https://doi.org/10.48550/arXiv.2112.02694.

¹⁷ L. Nasvytis, K. Sandbrink, J. Foerster, T. Franzmeyer, and C.S. de Witt, 2024, “Rethinking Out-of-Distribution Detection for Reinforcement Learning: Advancing Methods for Evaluation and Detection,” arxiv, https://doi.org/10.48550/arXiv.2404.07099.

¹⁸ For code, see LinasNas, “DEXTER,” https://github.com.mcas-gov.us/LinasNas/DEXTER, accessed October 16, 2024.

constrained environments.^19,20,21,22 While it is true that federated learning is already being researched from a communication network research perspective in the resilient and adaptive communication networks core competency, coverage of federated learning as an ML research topic with a more applied and empirical focus similar to other projects within the universal computation area would seem warranted. Indeed, existing ML methods and workflows (especially those related to handling training data) do not always directly and cleanly translate to federated learning scenarios.^23,24 This is an area that is not only ripe for research, but directly applicable to existing ARL application areas of interest like RAMITE. There are also opportunities at the intersection of computational mathematics and federated learning.^25,26,27,28

Furthermore, given the dynamic and uncertain nature of the problems that require algorithms, as well as the topics presented, a concerted effort in stochastic processes, probability, and data assimilation would benefit ARL. This would support current topics of interest like RL edge computing, and federated learning. Additionally, investment in multimodal and multi-fidelity learning capabilities will benefit the competency’s need to process information that may be compromised (in quality and/or quantity), as well as information coming from various sensing mechanisms (including audio, text, and video).^29,30,31,32

Finally, large language models (LLMs) are an emerging research direction with applications relevant to this core competency and present a ripe opportunity for ARL. Moreover, researchers across all three core competencies expressed an interest in pursuing research in areas such as LLM inference at the

___________________

¹⁹ F. Wu, C. Dong, Y. Qu, H. Sun, L. Zhang, and Q. Wu, 2022, “CIOFL: Collaborative Inference-Based Online Federated Learning for UAV Object Detection,” IEEE 19th International Conference on Mobile Ad Hoc and Smart Systems (MASS), https://doi.org/10.1109/MASS56207.2022.00043.

²⁰ C. Zhang, X. Liu, A. Yao, J. Bai, C. Dong, S. Pal, and F. Jiang, 2024, “Fed4UL: A Cloud–Edge–End Collaborative Federated Learning Framework for Addressing the Non-IID Data Issue in UAV Logistics,” Drones 8:312.

²¹ S. Gupta, K. Ahuja, M. Havaei, N. Chatterjee, and Y. Bengio, 2022, “FL Games: A Federated Learning Framework for Distribution Shifts,” arxiv, https://doi.org/10.48550/arXiv.2211.00184.

²² M. Du, M. Zhang, Y. Pu, K. Xu, S. Ji, and Q. Yin, 2024, “The Risk of Federated Learning to Skew Fine-Tuning Features and Underperform Out-of-Distribution Robustness,” arxiv, https://doi.org/10.48550/arXiv.2401.14027.

²³ F. Wu, C. Dong, Y. Qu, H. Sun, L. Zhang, and Q. Wu. 2022. “CIOFL: Collaborative Inference-Based Online Federated Learning for UAV Object Detection.” IEEE 19th International Conference on Mobile Ad Hoc and Smart Systems (MASS), https://doi.org/10.1109/MASS56207.2022.00043.

²⁴ J.-H. Ahn, K. Kim, J. Koh, and Q. Li, 2022, “Federated Active Learning (F-AL): An Efficient Annotation Strategy for Federated Learning,” arxiv, https://doi.org/10.48550/arXiv.2202.00195.

²⁵ L.-Y. Wei, Z. Yu, and D.-X. Zhou, 2023, “Federated Learning for Minimizing Nonsmooth Convex Loss Functions,” Mathematical Foundations of Computing 6:753–770.

²⁶ H. Zhao, K. Burlachenko, Z. Li, and P. Richtárik, 2024, “Faster Rates for Compressed Federated Learning with Client-Variance Reduction,” SIAM Journal on Mathematics of Data Science 6, https://doi.org/10.1137/23M155382.

²⁷ J. Ding, E. Tramel, A.K. Sahu, S. Wu, S. Avestimehr, and T. Zhang, 2022, “Federated Learning Challenges and Opportunities: An Outlook,” CASSP 2022–2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), https://doi.org/10.1109/ICASSP43922.2022.9746925.

²⁸ P. Yang, H. Zhang, F. Gao, Y. Xu, and Z. Jin, 2023, “Multi-Player Evolutionary Game of Federated Learning Incentive Mechanism Based on System Dynamics,” Neurocomputing 557:126739.

²⁹ Y. Qu, J. Nathaniel, S. Li, and P. Gentine, 2024, “Deep Generative Data Assimilation in Multimodal Setting,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops.

³⁰ F. Bao, Z. Zhang, and G. Zhang, 2024, “A Score-Based Filter for Nonlinear Data Assimilation,” Journal of Computational Physics 514:113207.

³¹ E. Walker, N. Trask, C. Martinez, K. Lee, J.A. Actor, S. Saha, T. Shilt, D. Vizoso, R. Dingreville, and B.L. Boyce, 2024, “Unsupervised Physics-Informed Disentanglement of Multimodal Data,” Foundations of Data Science 7(1):418–445, https://doi.org/10.3934/fods.2024019.

³² F. Stadtmann, E.R. Furevik, A. Rasheed, and T. Kvamsdal, 2024, “Physics-Guided Federated Learning as an Enabler for Digital Twins,” Expert Systems with Applications 258:125169.

edge.³³ One promising new research area that connects universal computation in the computational methods for modeling and learning core competency with the resilient and adaptive communication networks core competency is the application of LLMs for wireless communication networks. For instance, Shao and colleagues (2024)³⁴ recently introduced several promising wireless intelligence use cases, including power allocation, spectrum sensing, and protocol understanding. Moreover, there is growing interest in employing LLMs in scientific computing in areas such as material science³⁵ and genomics.³⁶

Opportunities Identified for Individual Projects

• Generalizing Out-of-Distribution with Deep Reinforcement Learning (Poster). Concerning this project, it is noteworthy that no existing RL models or combination of models were found to be able to adequately handle distribution shifts. This is not a surprising result, as adapting or generalizing to out-of-distribution examples without re-training is a challenging problem.³⁷ Thus, this project might be considered comparatively more at risk of not meeting its objectives than others. Given the degree to which ARL’s research focus is on computing and communication to the tactical edge, it should be noted that federated learning has been cited as an approach to addressing distribution shifts in the field.³⁸
• Projects: Adaptive Inference Framework of Video Object Detection on Edge Devices for Dynamic Environments and Methods for Accelerated Inference of Neural Networks on Resource Constrained Devices. The presentations on adaptive and accelerated inference demonstrated sufficient understanding of the underlying science and technology developed in the field. This is a rapidly evolving area of research, and staying current with the limited staff size working on the problem presents a challenge. In addition to computer vision forums (e.g., the Conference on Computer Vision and Pattern Recognition [CVPR] and the International Conference on Computer Vision [ICCV]), which consider the joint algorithm-platform problem from the algorithm perspective, researchers working in these areas are encouraged to also look into leading embedded systems venues (e.g., the Association for Computing Machinery [ACM]/Institute of Electrical and Electronics Engineers [IEEE] Design Automation Conference [DAC], the IEEE/ACM Design, Automation and Test in Europe [DATE] Conference, and the ACM Transactions on Embedded Computing Systems [ECS]), where they approach the same problem from the platform viewpoint.

___________________

³³ See, for example, Z. Yu, Z. Wang, Y. Li, H. You, R. Gao, X. Zhou, S.R. Bommu, Y.K. Zhano, and Y.C. Lin, 2024, “EDGE-LLM: Enabling Efficient Large Language Model Adaptation on Edge Devices via Layerwise Unified Compression and Adaptive Layer Tuning and Voting,” arxiv, https://doi.org/10.48550/arXiv.2406.15758.

³⁴ J. Shao, J. Tong, Q. Wu, W. Gao, Z. Li, Z. Lin, and J. Zhang, 2024, “WirelessLLM: Empowering Large Language Models Towards Wireless Intelligence,” arxiv, 2405.17053v2.

³⁵ See, for example, Y. Hu and M.J. Buehler, 2023, “Deep Language Models for Interpretative and Predictive Materials Science,” APL Machine Learning 1:010901.

³⁶ M. Zvyagin, A. Brace, K. Hippe, Y. Deng, B. Zhang, C.O. Bohorquez, A. Clyde, et al. 2022. “GenSLMs: Genome-Scale Language Models Reveal SARS-CoV-2 Evolutionary Dynamics,” bioRxiv, https://doi.org/10.1101/2022.10.10.511571.

³⁷ J. Liu, Z. Shen, Y. He, X. Zhang, R. Xu, H. Yu, and P. Cui, 2021, “Towards Out-Of-Distribution Generalization: A Survey,” avXiv, https://doi.org/10.48550/arXiv.2108.13624.

³⁸ F. Wu, C. Dong, Y. Qu, H. Sun, L. Zhang, and Q. Wu, 2022, “CIOFL: Collaborative Inference-Based Online Federated Learning for UAV Object Detection,” IEEE 19th International Conference on Mobile Ad Hoc and Smart Systems (MASS), https://doi.org/10.1109/MASS56207.2022.00043.

Cyber Defense and Cybersecurity Core Competency

Achievements and Advancements

During the August 6–8, 2024, review at the Adelphi Laboratory Center, four presentations and five posters were presented by ARL to represent extramural and intramural research being performed by the cyber defense and cybersecurity core competency. Overall, the extramural project portfolio was appropriate, to the extent presented in the review. The work was evaluated in three focus areas: trusted autonomy, defensive deception, and adversarial clean-label defense for ML. Leading researchers and institutions engaged in the research were recognized. The results are being published in notable venues. The work addresses core competency needs and the PSQs, which are reasonable, good questions.

The intramural work on trusted autonomy (the Autonomous Intelligent Cyber-Defense Agents for Army Vehicle Platforms project) also appears reasonable. The work in trusted autonomy generally reflects a broad understanding of the underlying science and technology. Some opportunities for enhancement of the impact of this work are discussed in the next section.

The research into autonomous intelligent cyber-defense agents for vehicles and vehicle cyber resilience was well structured. The effort has a strong scientific lead, and the team has a reasonable future research plan. This is an active research area outside ARL, and the research being done at ARL is on par with that of external research. Notably, the team has some current and future collaborations already planned, and they are encouraged to stay engaged with their scientific community around these topics.

Research Portfolio Opportunities

The ARL extramural portfolio of projects is generally strong. There is appropriate expertise of performers and a good balance of collaborations and disciplines. Still, ARL intramural researchers and managers could consider enhancing connections to relevant partners in areas such as cybersecurity, information assurance, and information manipulation. Connections to research-focused groups in industry, such as Microsoft Research, Google Research (e.g., their team focused on adversarial AI), and Amazon’s AI team, would be beneficial. Several universities have established research centers focused on these topics as well. The Computing Research Association’s (CRA’s) CRA-I committee could potentially facilitate that type of interaction. Connections could also be made to relevant federally funded research and development centers (FFRDCs). Additionally, connections with higher educational institutions, including those that have Department of Homeland Security’s (DHS’s) National Security Agency (NSA) Center of Academic Excellence-Research (CAE-R) designations, could also be established. These connections could increase awareness of ARL’s extramural opportunities and perhaps help increase interest in internships and permanent positions at ARL.

Furthermore, while it is not recommended that the extramural research portfolio be refocused on AI/ML, ARL’s extramural managers could ensure that it has appropriate coverage of those topics, particularly in the context of information integrity research areas.

The composition of the intramural cyber defense and cybersecurity core competency portfolio of projects appears to be narrow, and ARL may want to consider taking a broader view of the cybersecurity life cycle to include elements of design, response, and recovery in addition to detection and defense. Consulting external documents such as the NIST NICE education model and the NIST Cybersecurity Framework may help to seed a broader view of the field.

The work in cyber deception was technically sound but lags behind what has been accomplished elsewhere. This work appears to be disjointed from what is already available commercially, and from what may be feasibly deployed in any real environment. As such, it seems unlikely there will be any significant impact from this work beyond some potential academic publications.

ARL was an early mover in defensive cyber deception, leading to pioneering research in that topic 12–15 years ago. That has led to the work in deceptive honeypot technologies presented during the review (the “Dynamic Honeypot Allocation” presentation). However, the commercial marketplace also

embraced deception, and there are many products and companies currently offering deception and honeypot technologies as commercial-off-the-shelf (COTS) and government-off-the-shelf products. Examples of these companies include Countercraft, Acalvio, Attivo, Illusive Networks, TrapX, Smokescreen, and others. The current ARL research in honeypots has been surpassed by some of these offerings and by other research efforts. It will be important for ARL researchers working in this domain to broaden their awareness of commercial products in order to stay current and avoid duplicative efforts.

Additionally, it should be noted that some of the honeypot work appeared to be shallow, with unduly constrained assumptions (e.g., that attackers would not know that deception was in use and that there were limited entry points into the network being protected). This limits any generalization of the results to environments such as wireless mesh networks.

Deception, as a defensive technique, is likely to present some opportunities for innovation, so the concern is not for research into the concept generally. Overall, this area of research needs to be critically examined to determine what unmet needs might be addressed by future research within ARL. Any such research will need to be structured with an understanding of what is available commercially and with realistic assumptions about operational contexts. ARL informally noted that they desired feedback on which projects and research areas they could double down on versus where they need to go in a different direction. Given the limited resources, novelty, and impact of this current work, ARL might want to consider pivoting its focus on research avenues that may bear more fruit.

Overall, the intramural projects appear unduly focused in a few areas. No efforts were presented that were directed to reconstitution, active defense, supply chain assurance, insider threat, or secure by design, as examples of other topic areas. Consideration needs to be given to the breadth of cybersecurity topics to identify potential research that may apply in an operational environment. This will require balancing available resources with overall needs.

Generally, across all competencies, ARL work might be better focused by explicitly defining measures of success, threat models, and assumptions. These may be defined for some projects, but they were not presented; when asked, some researchers were unable to explicate these concisely. Certainly, basic research (6.1) needs to not be unduly constrained to enable exploration and innovation, but it also should not be unbounded when resources are limited and so an enhanced focus on these three elements could be important.

It should be noted that Sandia National Laboratories are beginning a new, multiyear effort in Digital Assurance for High Consequence Computing Systems. Exploring opportunities for synergistic collaboration could enhance research in ARL’s overall portfolio in information assurance. There appears to be a stub website at https://www.sandia.gov/research/digital-assurance-for-high-consequencesystems.³⁹ In addition, NIST is coordinating several activities related to trustworthy AI/ML development.⁴⁰

Opportunities Identified for Individual Projects

• Adversarial Clean Label Defenses for Cyber project. This project’s work in clean label defense is comparable to work being done on this general topic elsewhere. It was not possible, however, to fully evaluate the overall likely trajectory of the project as the results presented did not include a confusion matrix and involved limited trials. The project showed good performance (in the 90+ percent range of correctness), but there appeared to be a significant number of incorrect identifications that might make the approach impractical in practice. This work appears to be in an early stage, so additional effort may lead to better results. The project would also benefit from an understanding of the impact of scale coupled

___________________

³⁹ Sandia National Laboratories, “Digital Assurance for High Consequence Systems,” https://www.sandia.gov/research/digital-assurance-for-high-consequence-systems, accessed October 21, 2024.

⁴⁰ National Institute of Standards and Technology, “Trustworthy and Responsible AI,” https://www.nist.gov/trustworthy-and-responsible-ai, accessed December 20, 2024.

• with response. For instance, even a 1 percent false positive rate on 1 million instances would result in human operators ignoring or disabling the systems; evaluation of 10,000 instances to correct labeling would not be acceptable. This work (and related work external to ARL) may eventually result in better performance; however, the project would benefit from a clearer definition of what success looks like. This is an area of active research in both academia and industry, so the researchers will want to keep current with that work. This project might tie into the trusted autonomy project (overarching vehicle cyber resilience project’s associated project name, Autonomous Intelligent Cyber-Defense Agents for Army Vehicle Platforms), if there is not already a connection.
• Autonomous Intelligent Cyber-Defense Agents for Army Vehicle Platforms project. The project team may need to broaden their approach to consider hybrid defenses: blending humans and/or static defenses into the agent-based approach they are exploring. This would help address problem areas where an agent-based approach might be limited or suffer significant error. There is also an opportunity to integrate reconstitution and fallback technologies as part of trusted autonomy—detection alone will not solve the autonomy problem that was presented as the motivation for this research.

  A more explicit definition of assumptions might be useful in scoping the work. For example, what is being assumed about access to the controller area network bus (CANbus)? What are the modes of failure in the supply chain? Overall, what is the threat model they are seeking to counter? If they have not already done so, the team could connect with NSA’s advanced research division (known as “NSA R”) and Information Assurance Directorate divisions, and DHS’s Cybersecurity and Infrastructure Security Agency, for a better understanding of what they have done on this general topic and for potential collaborations. Also, they may want to connect with major vendors in this area, such as General Motors, if those connections do not already exist. There are also opportunities to potentially engage with the Defense Advanced Research Projects Agency (DARPA) on some of the cyber and AI challenges.

Resilient and Adaptive Communication Networks Core Competency

Accomplishments and Advancements

The research efforts within the resilient and adaptive communication networks core competency are addressing contemporary challenges across a wide breadth of fields, with a few notable gaps; using current and forward-looking analysis techniques, with a few deficiencies; and developing and making use of good and appropriate laboratory infrastructure, with a few deficiencies. A conversation about the strengths and challenges of different areas of the core competency as well as suggested pathways to the remediation of identified deficiencies are discussed below.

Leadership in key directions that include quantum frequency conversion and connecting different modalities has been demonstrated, and there are no major risks of the core competency not meeting its objectives. There is a solid theme in the portfolio in exploring the hardware layer in the classical and quantum realm and the efficient exploitation of resources at the edge (computing/networking/sensing and their mixtures, etc.). The core competency as a whole demonstrates a broad understanding of classical networking and quantum networking with its PSQs. Indeed, the PSQs are excellent and present no gaps across the fields. For the classical networking research thrust, the high-level directions set by the PSQs are good and are addressing many important current issues in the field. The specific programs and methodologies, as presented, that were chosen to address the PSQs were well thought out overall.

The research methods and methodologies used by researchers in the core competency are sound, well thought out, and complementary. In the quantum networking efforts, the combination of efficient analysis framework, novel device development, and collaborative system-level demonstrations presents a blend of capabilities that is highly desirable and rare. The classical coherent modulation and

demodulations, however, will need to be more focused on quantum networks and not on classical optical networks, which are very different. In the network resource allocation efforts, the Lyapunov drift-based and virtual queue approaches are built on classical approaches in the literature, which provide theoretical performance guarantee and strike a good balance between the optimality and constraint violation tradeoffs.

The distributed and federated ML efforts appropriately use the widely adopted FedAvg algorithm as the foundation to further investigate and develop joint networking-computing control and optimization algorithms. It should be noted that while AI and ML are useful tools, classical statistical and math programming techniques should not be abandoned. It is important to keep math programming research alive at ARL and search for fast network adaptation algorithms. Fast algorithms are needed for dynamic network applications. ML will not be useful if a situation develops that is very different from the range of data that was used in the model training process. Classical methods of detection need to be used in conjunction. This is discussed in further detail below.

The core competency’s portfolio was found to be balanced, and it would not be beneficial to drastically move into an entirely new direction in any area. That said, the classical communication networking thrust identified important opportunities to bolster research efforts, especially those focused on routing and network architectures. The current quantum networking efforts are excellent but may benefit from the inclusion of distributed quantum enhanced sensors and expanding the distributed and networked quantum processor efforts and collaborations. A deeper conversation about these opportunities is provided below.

Research Portfolio Opportunities

The structure of the efforts within the core competency is split between classical communication networking and quantum communication networking, and as much as the areas are structured to collaborate and share knowledge, there are inherent divisions as a result of the general differences between them in underlying physics principles, potential capabilities, and technological maturity. As a result, the assessment below broadly treats the classical efforts as one thrust and the quantum efforts as another.

Classical Communication Networking Research Thrust

The classical communication and networking research thrust, the efforts are progressing and making interesting contributions. However, the scientific quality of the research, compared to other leading peer institutions, is hampered by not addressing the fundamental challenge of network routing to a satisfactory level. In addition, the primary limitation of the classical communication and networking efforts at the program level, are that the program does not substantially address the challenges of multi-hop network architectures and routing in complex topologies with heterogeneous systems and in potentially adversarial scenarios.

It is vitally important to note that the architecture of infra-structureless wireless has never arrived at a point that can exceed 12–15 nodes. It is mostly a Layer 3 routing layer problem in conjunction with a Layer 1 physical layer connectivity problem. In addition, two major issues have been obstacles for more than 20 years. The central issue is when there are many nodes (>15), “whisper to the nearest neighbor” algorithms do not scale in two ways. First, most nodes are passing pass-through traffic from others and greatly diminish their ability to serve local traffic. The scaling factor for the fraction of the maximum node rate that can be used to serve local traffic is 1/$\sqrt{N}$logN where N is the total number of nodes. For any network bigger than ~15 nodes, the actual local traffic source and sink rate of a node is less than 5–10 percent of the node’s maximum physical communication rate. Second, whatever fraction of the communication rate that remains from the hop-by-hop routing must have the capacity to support network topology discovery—which is especially important in mobile scenarios—via a link-state routing protocol,

usually via flood routing. This will consume the capacity left by the first point and render the network unable to find an adequate spanning tree for routing.

The above points are fundamental physics and math limitations and have been supported by extensive experiments funded by the Department of Defense (DoD). There are a few possible solutions proposed by the DARPA Strategic Technology Office (STO) between 2010–2020 and in some open academic literature, but they have never been analyzed carefully and holistically in a proper setting. It behooves ARL to quickly study and settle on scenario-appropriate architectures for its research teams to have correct expectations of the would-be network capacity and behavior for its use scenario. This is vital to ensure that all the good research done at the laboratory, created on top of the network service, is not built on shaky ground.

On a different note, the absence of a digital twin effort at ARL for testing algorithms and blue and red team style defense and attack on network is a notable opportunity for improvement. High-fidelity digital twins are expensive but vital to certify many architecture constructs, modifications, and updates that cannot be proven.

The work on resilient and adaptive communications networks is generally well formulated and clearly presented. The use of ML and classical optimization techniques was described. ML alone will not be sufficient, and mathematical optimization techniques are also important. ARL’s decision to terminate math programming research is potentially harmful because there is a pressing need for fast (often suboptimal) algorithms for fast online applications. Optimum math programming provides very important benchmarks for fast suboptimal algorithms and, at times, math programming techniques must be used when ML models reach their limits. For example, fast changing events in a battlefield environment are unique scenarios that cannot simply be represented as uncertainties in the ML model. ML needs large amounts of data for training, and in some dynamic, topology-changing environments, as seen in battlefield environments or sudden adversarial attack scenarios, there will not be enough time for this training-based learning. In these situations, classical non-ML methods will need to be used. Additionally, in general, there needs to be a separate assessment of the complexity of all the algorithms in theoretical and quantitative form, in addition to the experimental observations.

In addition, AI opens up new attack surfaces and its vulnerability assessment and prevention needs to be a priority for ARL to study. Commercial companies are looking into all forms of attacks, including poisoning of data, manipulation of time stamps of inputs, etc. Thus far, mitigation techniques use guardrails, but better solutions need to be developed. There are several important efforts elsewhere in the government (e.g., in DoD’s Chief Digital and Artificial Intelligence Office [CDAO] and NSA’s Artificial Intelligence Security Center and R6 directorate). ARL could reach out to these entities to get updates of the latest understanding on the vulnerabilities and gaps that current mitigation techniques have not been able to address. One potential way to achieve the democratization of AI could be through a mixture of expert models and federated agents. See, for example, the work of Azalia Mirhoseini at Stanford University (her joint work with DeepMind).”⁴¹

Efforts by the Distributed Machine Learning at the Edge project; the Adaptive, Networked Analytics in Resource Constrained Environments project; and the Routing Strategies for Tightly Integrated Heterogeneous Networks project are making good and interesting progress, as evidenced by some of the new convergence rate results that demonstrate linear convergence speedup under arbitrary client selections. However, based on the current research results and progress, these three research efforts may be further improved by considering more recent and advanced distributed and federated learning paradigms, since the current federated learning architecture as presented remains quite standard. Real-time retraining of ML models deployed in the field is vital as situations change and new sensor data present themselves. The formulation of this research area is generally sound, with a notable underrepresentation of wireless network capacity and connectivity constraints. This is another manifestation of the absence of a solution for the wireless mobile network Layer 3 routing and achievable

___________________

⁴¹ B. Brown, J. Juravsky, R. Ehrlich, R. Clark, Q.V. Le, C. Ré, and A. Mirhoseini, 2024, “Large Language Monkeys: Scaling Inference Compute with Repeated Sampling,” arxiv, https://doi.org/10.48550/arXiv.2407.21787.

capacity for Layer 1 and Layer 2. These limitations will be a hard and first-order constraint on what this federated ML can do.

Another promising new research area in distributed and federated ML at the edge is to investigate communication-, data-, and parameter-efficient pre-training and fine-tuning pipelines for large foundation models in edge network settings. The astonishing success of LLMs and generative AI (GenAI) and the dominance of big tech companies in this domain have led to a compelling need to “democratize AI.” One promising approach in democratizing AI is to enable the pretraining and finetuning of LLMs that support GenAI using a vast amount of edge devices with massive parallelisms. However, realizing this goal is highly non-trivial and calls for fundamental breakthroughs in computing and networking. These new techniques also come with new attack surfaces. Thus, in parallel the security aspect of AI would need to be addressed.

The security of ingested data in the field and updating neural nets is extremely important, with a large body of available literature from both partners and adversaries on how to manipulate and poison such updates. An important but generally not well-addressed frontier is determining how to forget old and irrelevant data. In a benign mobile situation, new information due to topology and traffic changes must be quickly incorporated, and old information must be decommissioned. Also, in an adversarial situation, poisoned data need to be purged. The field has yet to find a graceful solution to this very important problem. Since all ingested data are used to train the weight parameters in specific instantiations, if sensed data are obsolete or bad, research must be done on how to negate their effects on weights. The fastest convergence in a dynamic environment is needed since situations can change quickly but the algorithms (e.g., neural nets with small number of nodes and trained by a small set of data) may become unstable and vulnerable with the slightest attacks. Classical statistical technique needs to be used to augment slow learning algorithms to sense the onset of attacks and react quickly. Such hybrid algorithms are not well researched, much less tested in realistic simulators or real-life experiments. A major suggestion for this effort is to incorporate a security aspect immediately. It may change some of the suggested paths forward.

The Adaptive, Networked Analytics in Resource Constrained Environments project effort has considerable overlap with the project on Distributed Machine Learning at the Edge, at least in name of the topics. Multi-hop wireless networks were explicitly mentioned, but there was no discussion on routing and capacity constraints. This is a distinct gap in the ARL research portfolio that will be important to address. This problem needs to be addressed first, or the rest of the efforts built on top of the network will have no assurance that the network can support the applications stated. Distributed analytics will require a common picture to emerge and receive agreement on eventually. It will be important to identify the choke points to arrive at a common picture. There is no mention of the database intended to be used, whether it is, for example, structured (e.g., structured query language) or unstructured (e.g., Hadoop) is not clear. If a structured database is used, there can be scaling problems and computation and search speed issues. If the database is unstructured, there is the possibility of putting the wrong weights on multiple reports from the same source and even having new updates negated by stale data. Perhaps there is a need for some form of heritage to track bad data repeatedly ingested or multiple reports of information from the same source. It may also be important for security to have some record of provenance of reports.

ML could also potentially fit into the core competency’s effort by allowing the development of new channel codes; however, it would be a new direction for the field. Indeed, channel codes use specific logical structures matched to specific physical channel and system limitations for easy decoding. ML usually has no underlying physical principles, and thus, using it for decoding is expected to result in significant decoding complexity, until proven otherwise.

The Army Research Office (ARO) Wireless Communications Program is making significant progress with a critical mass of efforts, doing field-leading work related to PSQs relevant to this core competency. Many small research programs under this umbrella program were shown as exemplars during the review. A significant fraction are good efforts that address parts of the big problem. Most of the principal investigators (PIs), however, were information theorists with little coverage of Layer 3 (routing) and up. Expertise in Layer 3 and up is sorely needed at ARL, and so, if ARL is not bringing on new collaboration projects that incorporate Layer 3 expertise, it may want to consider doing so. As

mentioned previously, a focus on routing multi-hop networks could be added as a research focus. All the excellent efforts at ARL will depend on adding such a focus. ARL will need to have a view of the Layer 3 problem first so that the range of other programs can use a common set of realistic assumptions.

The ARO Wireless Communications Program could also benefit from a coherent focus on integrating work on routing challenges, congestion control, scheduling, and restoration from failures into the program. This is an area where all the network layers must work together including the application. Without a holistic view and a program that reflects that, there is little chance of constructing a viable network architecture.

It should be noted that the Exploiting Novel Physical Layer Technologies for Low Probability of Detection Networking project effort and the Routing Strategies for Tightly Integrated Heterogeneous Networks project effort are addressing very important questions, but they are behind the state of the art compared to, for example, Office of Naval Research (ONR) efforts led by Santanu Das. The LPD work has been done with the assumption that crypto-driven techniques are not used. It is very limiting to not be able to use cryptography. Nonetheless, classical techniques such as pseudo-noise sequence signaling can be used. A public key can easily synchronize the transmitter and receiver. It does not have to be chosen at random, and the codebook has to be somehow sent to the receiver. If the sequence used is less than ~logN where N is the period of the sequence generator, the chance of the adversary figuring out the key is very small. It is not advisable to use a sequence for longer than logN before refreshing, because of known classical attacks. It would be useful if the research unit contacts NSA Research’s R2 group to get further insight. With the proper pseudo-random sequence cover, most of the techniques proposed become nice to have but not necessary. The problem statement of the Routing Strategies for Tightly Integrated Heterogeneous Networks project effort is a very important research direction because of the attention to heterogeneous networking. It will be important to nurture this area in the future with the following added components:

• The routing algorithm in the current version is confined to static networks. In the field, mobile networks are extremely dynamic, and the speed of reconfiguration stresses the Layer 3 protocols and underlying Layer 1 and 2 connectivity and capacities.
• The physical connectivity of Layer 1 is assumed, but therein lie the problems both of scalability and Link State Protocols burden on the throughput. There must be a baseline algorithm to establish the everchanging and fluid physical connection topology before the investigation of routing algorithms over an assumed topology.
• The presentation assumes static and centralized control. This is not survivable in the field, and distributed control of the network will need to be explored (stated as a future goal of the program).

The above components need to be addressed, or the set of excellent routing algorithms and applications will be built on quicksand. The Dijkstra algorithm—considered at ARL for routing—as well as the Bellman-Ford algorithm and the Floyd-Warshall algorithm—which are appropriate for secure routing—are only good for optimizing cost functions made up of additive metrics such as latency. However, relevant applications on top present nonlinear cost functions (this is an area where math programming is needed) on networks. As a result, a more insightful model of the routing problem, analyzed in conjunction with applications, needs to be developed. Different applications will impose a dynamic offered traffic to the network, which will affect its capacity, flows, and the routes chosen. On the issues of LPD and low probability of intercept (LPI), it will be very beneficial to connect with the Laboratory of Telecommunication Science (R4) and Directorate of Science and Technology at the Central Intelligence Agency (DS&T), which has expertise in these areas.

Opportunities Identified for Individual Projects

This section defines opportunities or expands on the conversation above for individual projects.

• Distributed Machine Learning at the Edge project. This effort is not yet on par with peers in the field in that it is not benchmarked to leading commercially available solutions, in particular, to SenseTime’s performance. For calibration, ARL could reach out to government peers such as CDAO, NSA’s Artificial Intelligence Security Center, and R6 directorate. The effort would also benefit from a focus on how to tag data heritage, discriminate good data from bad data, and unlearn and purge bad, repeated, or corrupted data, as well as extensions to federated and distributed multi-level learning, federated multi-objective learning, and federated learning with non-conventional assumptions. The project would also benefit from expanding on the current quantum networking efforts. All efforts to touch base and compare notes with government peers and academia will be greatly beneficial.
• Exploiting Novel Physical Layer Technologies for Low Probability of Detection Networking project. Continuing the analysis of this project from the previous section, this project is not yet on par with peers in the field because it is not addressing critical vulnerabilities in the effort to codebook capture. This can be remediated by using a simple cryptosystem such as a public key to generate pseudo-random sequences. The effort would benefit from a focus on how to address robustness and vulnerabilities to codebook capture as well as routing challenges. It would be useful if the research unit contacts NSA Research’s R2 group and DS&T to gain insight.
• Resource Allocation of the Network, Compute and Analytics project. This project is looking at important and very hard problems that are longstanding problems in the networking research community. As a result, the space for making significant advancements may be limited. However, ARL’s effort in this area is still making interesting contributions, and its progress is in line with peer organizations, as demonstrated by some of the new networking-computing co-design problems that jointly optimize the combined learning convergence and communication time bounds, which have been published in networking conferences.

  Opportunities to increase the impact of this project effort are to incorporate a focus on routing, and on how to perform fast computation in the field that is inherently scalable to many node architectures. It is recommended that the more topic-appropriate and impactful conferences in this area (e.g., IEEE’s International Conference on Computer Communications [INFOCOM], ACM’s Special Interest Group on Data Communications conference [ACM SIGCOMM], and ACM’s Mobile Ad Hoc Networking and Computing [MobiHoc] Conference) be targeted for presentation of results. There are a few possible solutions that were proposed at DARPA STO between 2010–2020 and in some open academic literature but never analyzed carefully and holistically in a proper setting. It behooves ARL to quickly study and settle on the right architecture and advise the rest of the research teams present during the review of the correct expectation of the would-be network capacity and behavior. This is vital to make sure that all the good research done at the laboratory created on top of the network service is not built on shaky grounds.

• Routing Strategies for Tightly Integrated Heterogeneous Networks project. This effort is making satisfactory and interesting progress; however, it is not yet on par with peers in the field because it has not addressed multi-hop networks with non-fully known, dynamic graphs. Additionally, most of the PIs are information theorists whose interests are primarily in the link Layer 1 and less to do with Layer 3 routing and upper layers. Adding additional expertise to this ARL effort is vital. Additional collaboration with the efforts in the other core competencies within the network, cyber, and computational sciences competency may be helpful.

Quantum Communications Research Thrust

Within the quantum communication and networking efforts, the scientific quality of the research, compared to other leading institutions, is of excellent quality. The high-level directions set by the PSQs are excellent, with no notable gaps. The specific programs and methodologies, as presented, that were chosen to address the PSQs, were overall well thought out. The main limitation of the quantum communication and networking effort is the lack of sufficient laboratory infrastructure to perform component device and system-level characterization, testing, and evaluation. This is further discussed in the “Facilities and Resources” section of this chapter.

The Photon Control in Quantum Networks project effort, the Ion-photon Entanglement at 780 nm for Hybrid Quantum Networking project effort, and the Towards Efficient Quantum State Characterization project effort stand out for their high-profile, leading work in thin film lithium niobate devices and deployed frequency conversion for interconnected quantum networking and network operation modeling. The work is of high value at both the applied and basic science levels, with multiple avenues to push science further. The work is essential for future photonics and quantum science-based technologies. This results in impactful publications and productive collaborations with established leaders in the field and doing important and needed work to help to cohere the emerging coalition of the Washington Metropolitan Quantum Network Research Consortium (DC-QNet) quantum networking stakeholders.

The assessment criteria asked for promising new areas that ARL could consider. A very promising quantum networking direction is to extend the focus to include how to attach quantum sensors and processors to the end nodes of quantum networks. Strategically, this could leverage and enhance ARL’s leading role in the DC-QNet. It could leverage existing analysis and hardware capabilities to determine optimum architectures for both scaling and connecting logical qubit processors, in line with and complementary to IonQ efforts and collaborations, as well as determine how to develop components for distributed sensor networks, in line with leading analysis work coming from researchers such as Quntao Zhuang at the University of Southern California.⁴² Higher-layer quantum network protocols (Level 3 and Level 4, the Transport Layer) are new horizons that would be important to have a place in the quantum portfolio.

Opportunities Identified for Individual Projects

• Photon Control in Quantum Networks project. For this project, the relationship between the device speed and precision—metrics relevant to both quantum and classical applications, versus cryogenic compatibility (a metric that is specific to most quantum applications)—is of paramount importance. Indeed, using a recirculating cavity can enhance modulation sensitivity in a cryogenic environment but greatly diminishes speeds (at 1 GBPS with current device and thus will not be used in classical coherent communications). Acacia (now a subsidiary of CISCO) and others have achieved 1.6 TBPS modulation in integrated silicon photonics and are moving on to 6.4 TBPS. The project team could connect with these organizations to familiarize themselves with the state of the art for classical applications. Thin films are an area that many other groups are pursuing. The research has merits even at the low speed of 1 GBPS if thin film techniques can substantially lower power consumption for quantum cryogenic applications. However, the coherent communication performance presented is significant under controlled I and Q channel rotations. Phase control is usually done with sensing and temperature tuning. In the absence of this functionality in hardware and software, the modulator is not ready for use. The overall power consumption must take into account this auxiliary hardware. It is not clear how the classical coherent communication

___________________

⁴² Z. Zhang and Q. Zhuang, 2021, “Distributed Quantum Sensing,” Quantum Science and Technology 6:043001.

• demonstration is relevant to quantum communication. For quantum networking, the error-correcting codes and transport layer protocols are very different from classical coherent communication. It will be fruitful to understand precisely what quantum networking needs before further investment in classical coherent communication experiments.
• Towards Efficient Quantum State Characterization project. In this effort, the general goal of utilizing distributed quantum systems to perform networking tasks is important. This includes synchronizing multiple quantum computers to do a single job. The narrow goal of characterizing the prepared quantum system for accuracy is a good starting point. While the stated long-term goal is “to verify the faithful distribution of entanglement needed to perform quantum networking tasks,”⁴³ the presented work targets a problem that is inherent in work with any quantum system—namely, how to efficiently and faithfully reconstruct a quantum state. One notable result of the effort is the utilization of ML with biased training sets (as opposed to randomized sets) that can be consequential for systems with limited resources. Another notable development is the discovery of a novel method for efficient parallelized Bayesian reconstruction, currently being tested at Oak Ridge National Laboratory, which takes advantage of high-performance computing (HPC) and is amenable to scaling. In each case, challenges, trade-offs, and use scenarios are well articulated, and the work may be consequential for the global scientific community. although it is true that the dimension of the Hilbert Space grows quickly in the number of identical copies of the same quantum state (the outer product Hilbert Space grows as 2^N, the sequential or parallel measurements necessary for characterization can be handled with less complexity of the general measurement in the combined space). The assumption that these are identical but independently prepared copies lend themselves to sequential measurements in each component space. However, the sequential measurement in general needs to decide what to measure, given the past results up to that time. This will greatly reduce the complexity of the measurement necessary. AI needs a lot of data, and it is not clear if this is the most efficient approach. Classical statistical inference can be a lot faster if used properly. There needs to be a comparison of AI and classical techniques (such as the Weak or Strong Laws of Large Numbers) on their convergence speeds and accuracies, especially if the prepared state is not a pure state but characterized by a density operator. Bayesian techniques potentially have better accuracies but have larger computation costs, and there was a lack of quantification of this trade-off in the presentation. The question is, Is the graph notional or is it based on analysis? The closest mention of Bayesian methods presented is using “mutation” algorithms to perform the estimation. Mutation techniques in classical applications tend to require a lot of computation and steps. Perhaps it is the same in the quantum case, thus other statistical methods will need to be explored.
• Ion-Photon Entanglement at 780 nm for Hybrid Quantum Networking project. Trapped ions are a mature platform with several envisioned applications, from positioning, navigation, and timing (PNT), to quantum sensing, quantum simulation, and computing. Lately, neutral atoms have emerged as another robust platform that enables these applications. Entanglement distribution is an anticipated core need for a quantum science-enabled ecosystem, whether geographically localized or not. The ion-photon entanglement and frequency conversion work at ARL explores one way to address this need. So far, this collaborative effort has resulted in remarkable state-of-the-art demonstrations, such as the first demonstration of a frequency-converted 780 nm photon entangled with a barium ion, or interfering photons emitted from ions and atoms. The importance of ARL’s achievement lies in the prospect that multiple quantum systems (trapped ions, rubidium atoms or Bose-Einstein condensates, and solid-state defects) with different properties and optimal use cases can be addressed at this wavelength,

___________________

⁴³ B. Kirby, 2024, “Towards Efficient Quantum State Characterization,” DEVCOM ARL presentation to the committee, August 6.

• allowing for entanglement distribution across different modalities with appropriately shaped photons, as targeted by ARL. Another point of significance is that at the longer photon wavelengths targeted by ARL, losses in materials decrease, allowing for both photonics processing and transmission through fiber and air at larger separations. This is an excellent project with very useful outcomes for synchronizing multiple quantum systems because photons are natural mobile carriers of quantum information or flying qubits. The current work is at 780 nm, along with a future extension to 1.5 micron as necessary for long-distance transmissions. The goal and future objectives of this effort are clearly defined and, to a considerable extent, ARL needs to incorporate the effort above on photon control and characterization as a larger effort. It may help the photon control effort focus much better on the important problems at hand.

  It should also be noted that the Photon Control in Quantum Networks project effort and the Ion-Photon Entanglement at 780 nm for Hybrid Quantum Networking project effort’s primary limitation is a lack of laboratory cryogenic optical infrastructure to perform onsite device characterization and use, leading to potential delay challenges. Better facilities and resources would improve these efforts and increase their pace. Current classical coherent modulation and demodulation efforts should transition immediately to a quantum network focus.

PORTFOLIO OF SCIENTIFIC EXPERTISE

Computational Methods for Modeling and Learning Core Competency

Overall, the ARL researchers within the computational methods for modeling and learning core competency were found to be competent, capable, and committed. The researchers (many of whom were more junior) should be commended for their enthusiasm and hard work. It appears that ARL will need to hire more staff whose research training and expertise is in computational applied mathematics. Only one presenter was identified as having been formally trained in applied mathematics, and this individual was a postdoctoral researcher, not a permanent staff member. Hiring more staff with expertise in applied and computational mathematics and developing critical mass in this area is important because this area is cross-cutting and fundamental to many things ARL wants to do, and the field is starting to explode with interesting scientific possibilities. Developing a coherent and state-of-the-art research portfolio in this area (e.g., computational mathematics, computational science, scientific ML, new algorithms, and methodologies for data-driven modeling of physical systems) demands more rigorous applied mathematical thinking. Re-tooling physicists, engineers, and computer scientists cannot address this gap. Expanding the ARL team with capable applied mathematicians will also help build needed critical mass for strategic project guidance, technical mentorship, and career mentorship of applied mathematics staff.

Additionally, it is not clear if the lack of a coherent vision described earlier in the chapter may, in part, reflect a career-level gap between management and the early-career researchers conducting most of the work. It was hard to estimate whether there are sufficient mid-career researchers that can steer research in addition to the high-level guidance provided by management. The addition of mid-career scientists may be able to better connect the high-level research goals in the competency with the presented PSQs. Mid-career colleagues can help to both provide a technically grounded strategic vision for projects and offer career mentoring to more junior colleagues. As good technical work does not happen in isolation, staff will need astute feedback and sounding boards from others with relevant technical expertise.

Additionally, stronger representation in computer science and a more concerted effort at ARL to participate in top-tier ML and embedded systems conferences (e.g., as mentioned, CVPR, ICCV, ACM/IEEE DAC, IEEE/ACM DATE, and the ACM Transactions on ECS) is also suggested. Such

engagements would not only better position ARL in keeping apprised of the latest research developments relevant to ongoing and future projects but also help with recruiting.

Finally, it should be noted that in addition to creating a stronger computational mathematics portfolio, recruiting experts in computational mathematics and computer science could create more capacity to collaborate with and distribute knowledge to other competencies at ARL. For example, with more staff in these areas, the network, cyber, and computational sciences competency could initiate compressed short courses for other competencies to raise awareness of ML, AI, and computational tools and techniques as a way of injecting AI/ML and computation into other scientific areas at ARL. In addition, more staff would allow the computational mathematics core competency to support other competency efforts by developing computational techniques and GenAI models specific to competency needs. This could be very helpful to the biological and biotechnology sciences competency and the photonics, electronics, and quantum sciences competency in particular.

Cyber Defense and Cybersecurity Core Competency

ARL personnel are competent and dedicated to their work. In particular, scientific leaders are among the personnel involved with the extramural projects. There was a good cross-section of personnel from institutions with different profiles and foci, and they were producing results appearing in well-recognized, peer-reviewed venues. This also speaks to the competency of the program managers who select and interact with those researchers. Several intramural researchers indicated that their research topics, or research-adjacent topics, were part of those researchers’ pursuit of a PhD/ScD. The ability and interest of master’s-level staff to pursue those advanced degrees is commendable.

There is a distinct shortage of intramural researchers with advanced degrees in computer science coupled with a demonstrated record of core computer science research. There is a notable difference in project framing and presentation involving topics that are primarily computing in nature between computer scientists versus electrical engineers, mathematicians, and physicists. Although physicists, mathematicians, and electrical engineers may have some background in elements of computing, there are disciplinary differences in approach. To that point, ARL needs to prioritize hiring or leveraging intramural computer scientists with research expertise in areas of ARL interest.

One measure of seniority and competency at a leading laboratory is external recognition; however, it is acknowledged that some ARL research cannot be publicized because of its security classification level. ARL personnel could still work toward such recognition, where possible, and help elevate peers to these distinctions. Senior researchers could strive to be senior members in the IEEE and/or distinguished members in the ACM; preferably, there could be several who are fellow rank within these organizations. It should also be noted that ACM now has the potential to recognize contributions in classified domains.

ARL could also engage with the computing community to better raise its profile and perhaps attract talent. Notably, extramural events with the ACM as contrasted with IEEE are more cybersecurity focused. Involvement with the non-profit CRA⁴⁴ might be helpful as well.

Efforts to bring in computing researchers from academia, industry, and other government laboratories via the Intergovernmental Personnel Act might also prove beneficial, leading to new insights, new collaborations, and possibly new hires. Other mechanisms, if available, such as temporary federal employees or appointments as visiting scientists and engineers, could also be explored. This approach might help to inject additional computer science expertise in the near term.

Finally, having a technical advisory board has proven to be beneficial to research entities in other contexts. A regular meeting of a cross-section of senior researchers from academia, industry, FFRDCs, and other government laboratories could provide insights, ideas, and guidance on a more frequent basis than the National Academies’ technical assessment board assessment reports. As in other contexts, the board members could be composed of senior researchers who hold appropriate clearances.

___________________

⁴⁴ See the Computing Research Association website at https//cra.org, accessed October 16, 2024.

Resilient and Adaptive Communication Networks Core Competency

The qualifications of the teams in the resilient and adaptive communication networks core competency are suitable and appropriate for the efforts. The teams, however, were comprised mainly of PhD-level researchers within about 10 years of experience, and the scope of efforts could support and benefit from a diversity of education and experience levels.

It should also be noted that there appear to be challenges in attracting and retaining intramural personnel, and more alacrity and latitude in hiring are needed to be competitive (i.e., to produce relevant work and/or have the relevant awareness level to eliminate danger of technological surprise). This is critically important to resolve for ARL to stay relevant.

The ARO Wireless Communications Program is making significant progress with a critical mass of efforts doing field leading work on the PSQs. Most of the PIs were information theorists with little coverage of Layer 3 (routing) and up. Expertise of Layer 3 and up is sorely needed at ARL, and so, if ARL is not bringing on new collaboration projects in this area, it may want to consider doing so.

FACILITIES AND RESOURCES

Computational Methods for Modeling and Learning Core Competency

With respect to the facilities supporting the computational methods for modeling and learning core competency, several of the projects—such as Adaptive LPD Radar Waveform Design with Generative Deep Learning, Scale-Bridging in Multiscale Models of Energetic Materials, General Neural Network Optimization Engine, and Deep Neural Network Sensitivity to Varied Radar Waveform Properties for Radar Classification and Domain Adaptation—mentioned use of ARL’s supercomputers. There did not appear to be any instances where ARL’s computing resources were limited in this core competency. It was gratifying to see ARL’s HPC environment being used effectively to support research objectives.

Cyber Defense and Cybersecurity Core Competency

Computer science is a laboratory science—while subdisciplinary aspects may be completed using only keyboard entry to a system, many areas of research require large storage, dedicated processing, virtual machine environments, emulators, reconfigurable network fabrics, customizable probes and monitoring, and possibly, human-subject experimental setups. This area of research requires exemplars of commercial and open-source artifacts to be used in experiments and comparisons.

The facilities supporting the cyber defense and cybersecurity core competency appeared to be insufficient to support its needs. While generalized computing and virtualization appears to be available to support most projects, what was missing was a reasonable collection of commercial and open-source products such as various honeypots and firewalls, Intrusion Detection System products, and Security Information and Event Management tools (to name a few). Having such tools would enable researchers to use existing platforms rather than build their own tools. It would also allow comparison with existing products and testing (some adversarial) and ensure currency with marketplace offerings.

Along similar lines, a more robust infrastructure needs to also house additional data sets, particularly for topics such as dynamic attacker/defender interactions and human decision theory models.

Notably, the hardware testbed from Toyota for the Portable Automotive Security Testbed with Adaptability (PASTA) project⁴⁵ appears to have an appropriate infrastructure. However, only a single instance of this, with vendor specificity, will likely be limiting as the project progresses. Having instances

___________________

⁴⁵ Chip1stop, “Portable Automotive Security Testbed with Adaptability PASTA,” https://www.chip1stop.com/sp/products/toyota-pasta_en, accessed October 16, 2024.

of the CANbus and processors for deployed platforms of interest available in a laboratory setting would provide better fidelity and flexibility for the various experiments in this project domain.

More generally, laboratory resources with actual instances of Internet of Things, unmanned aerial vehicles, and AI engines (as examples) need to be available and supported within the intramural work at ARL, both for experimental work in future projects, and to ensure that a core competency familiarity is maintained.

Computer scientists will note the presence or absence of laboratory resources. Thus, having a more robust laboratory would likely be an asset in hiring and retention; conversely, a meager or outdated set of resources will be viewed negatively and may impact ARL’s ability to recruit top-notch experts.

Resilient and Adaptive Communication Networks Core Competency

The onsite facilities and resources for classical networking are appropriate and are well matched to effort goals and staff expertise. The lack of a “digital twin” to test out new network protocols and applications riding on top of the network is a conspicuous void that needs to be filled. The onsite facilities and resources for quantum networking are appropriate but limited in a way that restricts what can be accomplished onsite. Fast-cycling cryogenic optical systems are becoming a staple in current state-of-the-art quantum networking laboratories and will greatly enhance ARL capabilities. As previously mentioned, both the Photon Control in Quantum Networks effort and the Ion-photon Entanglement at 780 nm for Hybrid Quantum Networking effort’s primary limitation is a lack of laboratory cryogenic optical infrastructure to perform onsite device characterization and use, leading to potential delay challenges. Better facilities and resources would improve these efforts and increase their pace.

Additionally, a more rapid method for purchasing miscellaneous components and consumables is needed to operate the laboratories in the manner that leading peer organizations do. There are challenges in procurement; procedures are antiquated and inadequate, in particular for the leading experimental work being done. For example, progress in a typical experiment may often require procurement of COTS components (general microwave components, small electronics, cables, fibers, optomechanics, solvents, cleaning supplies, etc.) available with same day shipping, and at times, delivery from vendors. It is not, in general, feasible or possible for researchers to plan for such needs. There seems to be no mechanism to utilize this purchasing flexibility at ARL, yet it is available to most other research institutions. In aggregate, this seriously hampers progress in fast-paced laboratories and is critically important to resolve for ARL to stay relevant.

The University of Maryland facilities and strong connection to ARL is a boon because of the tight integration with leading academic institutions and access to talent including foreign nationals. Connections to other communities and services (e.g., Naval Research Laboratory, Laboratory for Physical Sciences, Air Force Research Laboratory) can be improved to avoid duplication of effort, to leverage expertise, and to disseminate results and products.

THEMES THAT CUT ACROSS COMPETENCIES

Both the computational methods for modeling and learning core competency and the resilient and adaptive communication networks core competency noted the importance of including a focus on federated AI techniques (e.g., traditional federated learning and emerging research such as federated agents) into their portfolios and bolstered the importance of mathematics/statistics. In addition, both core competencies mention diversifying staff in terms of where they are in their careers (e.g., junior, mid-level), as this diversity will allow for more mentoring and possibly new scientific strategies toward achieving core competency goals.

Additionally, in both the cyber defense and cybersecurity and resilient and adaptive communication networks core competencies, a lack of facility resources is severely limiting the intramural researchers, and their outputs and may be affecting the intramural teams’ ability to recruit

needed expertise. This chapter outlines which resources are needed to build a robust intramural laboratory.

This chapter points out project-specific instances in both cyber defense and cybersecurity and resilient and adaptive communication networks core competencies where better knowledge of commercial capabilities may help avoid duplicative efforts or more quickly advance project goals. Thus, ARL will want to ensure that it keeps a closer eye on developments in commercial capabilities. Finally, enhanced engagement at conferences either through more attendance or more presentations was mentioned for all three competencies. Specific suggestions on relevant conferences are found within the chapter.