AFRL Tech Reports

This report details our approach to combining dynamic, distributed constraint reasoning with machine learning techniques and adaptive response strategies. By combining these technologies, we built a system that can 1) develop robust, adaptable mission plans 2) exploit knowledge learned through prior interactions with our adversary, and 3) autonomously and dynamically alter its behavior during mission execution to improve the likelihood of a successful outcome. This system has been thoroughly tested in the ATE2 and ATE3 simulators that were provided by AFRL/RI against four increasingly difficult milestones.

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1074604

The project started on Aug 18, 2016. In the 24 months of executing the project, the team has conducted basic research on data provenance architecture for cloud using block chain, surveyed the vulnerabilities in block chain, in-depth analysis of the block discarding attack, developed a Proof-of-Stake consensus protocol for cloud based blockchain, architecture for secure Battlefield IoT, cyber supply chain provenance, integration of software guard extensions on distributed ledgers for increased privacy.

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1068732

The DiRAM4 memory permits fast random access of DRAM via 64 bidirectional memory ports. A memory controller was designed and verified that is suited for interfacing a multicore CPU to this unique DRAM.

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1066447

This report summarizes the results of Project Clio. The goal of the Clio effort was to develop technologies that could help reduce the high cost of software by bringing more automation into development and maintenance tasks. Consistent with the overall goals of the Defense Advanced Research Program Agency (DARPA) Mining and Understanding Software Enclaves (MUSE) program, a central feature of our effort was to leverage data about existing software in order to enable capabilities that were previously not available.

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1069460

Cyber physical systems (CPS) form a ubiquitous, networked computing substrate, which is increasingly essential to our nations civilian and military infrastructure. These systems must be highly resilient to adversaries, perform mission critical functions despite known/unknown vulnerabilities, and protect and repair themselves during or after operational failures and cyber-attacks. We believe that an automated CPS repair approach that can prevent failures of related, mission-critical systems is a necessary component to support the resiliency and survivability of our nation's infrastructure. We integrated and evaluated techniques to cooperatively eliminate certain security vulnerabilities in CPS, to repair certain general classes of such systems, and to increase the confidence of human operators in the trustworthiness of those repairs and the subsequent system behavior. We worked with a Government-provided Red Team to demonstrate and validate our approach on embedded platforms, including an autonomous rover vehicle.

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1066449

The goal of the Space/Time Analysis for Cybersecurity (STAC) program is to develop new analysis techniques and tools for identifying vulnerabilities related to the space and time resource usage behavior of algorithms, including vulnerabilities to algorithmic complexity and side channel attacks. As an adversarial challenge team, BBN develops challenge applications that illuminate the behavior of the R and D team tools, allowing not only measurement of their performance, but also diagnosis of their behavior.

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1078527

Final report on the STAC project, Space/Time Analysis for Cybersecurity. Report covers performance of the Two Six Labs BASIC team, the Control team. Includes information on workflows.

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1081051

This research analyzed, evaluated, and characterized the computing and energy requirements for next-generation autonomous systems of significance to DoDs mission-essential tasks and thus investigated the effective and efficient computational intelligence approaches capable of supporting desired autonomy. This assessment will help determine the processing flow of an autonomous system from the cognitive perspectives, as well as the desired performance and energy requirements from the computing perspectives. Specifically, this study first outlined the necessary cognitive primitives and processing flow for a flexible autonomous system capable of real-time problem solving. Then, it focused on the autonomous target tracking problem and explored multiple computational intelligence methods, including artificial neural network (ANN), reservoir computing (RC), and deep learning (DL) architectures, to achieve the desired autonomy. Third, it investigated the computational characteristics of those intelligence models, assessed the performance metrics in terms of accuracy, speed, and energy consumption, characterized performance and energy requirements according to the scope of the problem, as well as identified the most suitable solutions fitting into the cognitive processing flow. Finally, it explored bio-inspired dynamic ensembles of reservoir networks for multiple pattern recognition, category learning driven classification network, and evolutionary adaptation of reservoir network optimization.

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1066668

The objective of this research work was to develop a ternary based public key exchange scheme that is addressable to replace, or complement, the existing Public Key Infrastructures (PKI). The Ternary Addressable Public Key Infrastructure (TAPKI) leveraged arrays of physical unclonable functions (PUFs) and heterogeneous ternary/binary computing systems. Public, and private key pairs are binary streams while the core of the T-PKA is based on ternary logic. The communication between parties can occurs over untrusted channels, by exchanging dynamically generated public keys, and using legacy binary codes. The proposed ternary environment largely enhanced entropy, creating an additional level of cyber-protection.

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1067826

This final technical report describes the result of a research project investigating economic issues involving the sharing of cybersecurity research data. It describes an effort to investigate data use and production in cybersecurity research publications from 2012-2016. Evidence is presented that researchers regularly use public data as input to research, but only rarely make created data publicly available. Additionally, it is shown that publications that do create datasets and make them publicly available are cited more often than those that do not. Additionally, utilization of the DHS IMPACT platform is investigated. Attributes of datasets are identified that are associated with greater demand from research consumers. Additionally, the value of sharing taken place on the platform is estimated to be approximately $663 million.

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1073243

Subject matter experts (SMEs) attempting to solve real-world analytic problems face several challenges due to the lack of applied mathematics, statistics, and machine learning skills that data scientists possess. The goal of our TA2 effort under the DARPA D3M Program was to span this gap by using novel methods and automation to enable SMEs to act as their own data scientists. Our effort was designed to fuse data- and knowledge-driven approaches to produce a virtual data scientist we call Eve. To translate domain-expert intent into formal representations of learning problems, we built a problem representation system that deterministically converts TA3 inputs into computer-interpretable mathematical expressions. To efficiently search for and compose the sequences of machine learning steps that comprise learning plans, we built a Monte Carlo Discrepancy Search approach that explores the vast space of possible plans through efficient modification and testing of prior related and/or successful plans. Further, we enriched these plans by incorporating data preparation models, treating data preprocessing functions as operators to be planned in-line with learning operators.

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1069846

In the role of adversarial challenger on the DARPA STAC program, CyberPoint developed many challenge programs designed to test the research tools being developed under the program for their ability to help analysts detect space-and time-related vulnerabilities in Java programs. In this document, we describe our approach to our Complexity and Side-Channel Adversarial Integrated Defects (CASCAID) effort, the vulnerabilities we developed, their efficacy, and what we learned in the process.

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1079266

The objective of this effort was to investigate the utility of hardware quantum annealing devices in two near-term applications: Circuit fault diagnosis; and machine learning. This report summarizes the technical work performed, results published, software developed, lessons learned, and future directions of study. Key accomplishments include: A new efficient algorithm for domain decomposition, allowing large optimization and sampling problems to be solved on small quantum hardware; A library of optimal Hamiltonians for common circuits; Software for rapid experimentation in quantum-assisted unsupervised Boltzmann machine training; and Training of several quantum hardware-native and non-native Boltzmann machines with state-of-the-art performance on standard benchmarks.

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1076914

This report discusses the design of three novel, cross-layer neighbor discovery schemes for directional wireless mesh networks, which have significantly lower protocol overhead and discovery latency. These schemes intelligently consider the collisions among the neighbor discovery messages of the neighboring nodes and use machine learning techniques. The performance of these schemes is evaluated for different network sizes, node densities, beamwidth and number of one-hop neighbors. Finally, implementation of directional neighbor discovery schemes in a hardware testbed is discussed.

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1069223

Domain-Specific Languages (DSLs) have the potential to make it dramatically easier to produce high-quality software in a timely and cost-effective manner. This potential has been difficult to realize, however, because DSLs require so much work to build. One way to mitigate this cost is to embed the DSL in a general-purpose host language. Embedded DSL programs are compiled down to the host language, where a complete suite of tools already exists. The problem with this strategy is that the host tools work at the host languages level of abstraction, essentially forcing the programmer to perform tasks such as debugging and profiling on the implementation of the DSL, which is often totally unrecognizable. The goal of this project was to develop a set of techniques and tools that make it easier for DSL designers to build first-class domain-specific languages, which come equipped with a full suite of support tools that operate at the level of abstraction of the domain. Users of these DSLs will get the productivity and code quality benefits of DSLs throughout the development lifecycle, from editing and compiling to debugging and profiling. Our approach uses an embedding strategy in order to continue to obtain other benefits from the host language, including general-purpose programming, access to existing libraries, and the possibility of employing multiple embedded DSLs within a single application.

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1069638

The study of dynamic behaviors of multiagent systems poses significant challenges due to the fact that complex system behaviors emerge as a result of their individual self-operating (sensing and actuating) capability, as well as of the interactions among agents. Some of those challenges include the comprehensive understanding of interaction mechanism among agents and the development of distributed resilient algorithms for detecting, designing, and controlling interaction dynamics and interaction topologies of multiagents. In this project, we studied fundamental issues from distributed estimation to distributed control of multiagent emergent behaviors. This multiple-year project leads to training of a number of graduate and undergraduate students and results in twenty-five research publications in the internationally refereed journals and conferences. Specifically, the highlights of the research outcomes include the following developments: 1) Designed a distributed estimation algorithm for the detection of certain global characteristic signals in multiagent systems, which can serve as feature indicators for the group collective behaviors. The event triggering mechanism for information transmission was further employed to reduce the communication load. 2)Designed a distributed sensor fusion algorithm for environmental monitoring by wireless sensor networks (WSNs) with limited communication.

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1065761

This effort made significant progress on inference for probabilistic programming. Probabilistic programming requires inference methods for approximating conditional distributions. Building on the framework of variational inference, this effort made this algorithm more efficient, more powerful, and more accurate. This effort developed new probabilistic models for economics, neuroscience, text analysis, population genetics, social network analysis, and recommendation systems. These methods were deployed in open-source software, on real-world programming systems and are currently in use by end-users of probabilistic programming. The work performed under this effort changed the landscape of approximate posterior inference, pushing forward the field of Bayesian machine learning and probabilistic programming.

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1067306

A tunable matching network design was designed for a K/V dual-band RF frontend. The tuning capability was separately demonstrated. Techniques for impedance tuning were investigated and achieved in hardware. The study determined the required range of impedance adaptation, and the most suitable method was identified to cover the desired impedance plane. A hardware prototype was developed consisting of the matching network, which was later integrated into a power amplifier. The benefit of the developed tunable network was experimentally demonstrated by comparing power amplifiers at K-band with and without tunable matching networks under varying load conditions using active load-pull system. K-band was specifically chosen for demonstration purposes due to the maximum frequency of the load-pull system. There were two integrated circuit fabrication runs to implement the successful demonstration.

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1070591

This report summarizes the activities enabled by an XDATA effort which includes a series of papers, the explosive growth of the Julia language, and a remarkable amount of software development. At the time of writing this report, there have been a near 4,000,000 downloads of the Julia Language, several textbooks authored by faculty worldwide based on the Julia language, and any number of classrooms using the Julia language. At the early start of the XDATA effort , Python was extremely popular, (as it remains today), but even large commercial companies such as Google are starting to understand Pythons shortcomings. At the same time, libraries written in Julia remain callable from other popular languages such as Python. A remarkable amount of documentation that is directly or indirectly a tributable to this work can be found on such pages as: 1. The Julia Lab web page at MIT: https://julia.mit.edu/; 2. The Julia Language webpage: https://julialang.org/; and 3. the authors MIT web page: http://math.mit.edu/edelman/

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1069148

The objective of this effort was to build affordable, manufacturable, low power, dynamic neuromorphic computing platform for handling spiky, highly variable information/data, as well as develop a low-power, hybrid memristor/CMOS neuron / synapse implementation for implementation of a hardware-based dynamic neural network array. The approach was to leverage hybrid CMOS/memristor encryption project to design, fabricate, and test memristor/CMOS neurons / synapses, and integrate them with existing FPGA implementations for full-scale dynamic neural network array demonstration. In addition, to make a module design kit library available for other circuits and architectures, thus making advancements in this effort useful to future designs that utilize hybrid CMOS/memristor technologies.

http://www.dtic.mil/get-tr-doc/pdf?AD=AD1079475