Software Engineering Institute

Grace A. Lewis

Principal Researcher

Grace A. Lewis is the lead for the SEI Tactical and AI-enabled Systems initiative and the principal investigator for several research projects in the field of Software Engineering for Machine Learning (SE4ML), including:

Continuum: Establishing the Practice of Integrated (T&E) for ML Capabilities

Delays in fielding of mission-critical systems in the DoD is a known issue, ranging from months to years, with problems found during developmental test (DT) and operational test (OT) as the most frequent cause for these delays. The growing use of machine learning (ML) capabilities in DoD systems expected to further increase these delays because of requirements for additional facets of testing plus the potential for failures during OT due to differences between operations and development test data and environments. Current practice for test and evaluation (T&E) of ML capabilities during development is largely limited to testing ML model properties, such as model performance (e.g., accuracy), without consideration of mission and system requirements, such as throughput, resource consumption, or robustness, leading to failures in model integration, deployment, and operations. The discovery of problems attributed to ML capabilities is especially problematic, not only because it requires to setup an additional OT event, but because fixing the problem might require additional data collection and retraining, further delaying system deployment. This is exacerbated by the fact that T&E organizations are segregated: OT organizations work independently from DT organizations which leads to inefficiencies; mission, system, and model developers doing contractor testing (CT) may not have access to appropriate mission and system requirements failing to adequately address the challenges and constraints of real-world operational environment. Motivation to solve these two problems has generated a push for Integrated T&E, an iterative approach to T&E in which system context and requirements, and intermediate test results, inform and are informed by T&E activities at all levels. The Continuum process and infrastructure is being implemented in a tool called MLTE.

Characterizing and Detecting Mismatch in ML-Enabled Systems

A problem with deployment of ML systems in production environments is that their development and operation involve three perspectives, with three different and often completely separate workflows and people: the data scientist builds the model; the software engineer integrates the model into a larger system; and then operations staff deploy, operate, and monitor the system. Because these perspectives operate separately and often speak different languages, there are opportunities for mismatch between the assumptions made by each perspective with respect to the elements of the ML-enabled system, and the actual guarantees provided by each element. We conducted a study with practitioners to identify and validate mismatches and their consequences. In parallel we conducted a multi-vocal literature study to identify best practices for software engineering of ML systems that could address the identified mismatches. The result of the gap analysis between the outcome of the two tasks is a set of machine-readable descriptors that contain attributes of system elements that need to be made explicit and shared between stakeholders to avoid mismatch. These descriptors can be used by system stakeholders for information awareness and evaluation activities; and by automated mismatch detectors at design time and run time for cases in which attributes lend themselves to automation. JSON descriptors for system elements are available on GitHub.

Automating Mismatch Detection and Testing in ML Systems

This project was a continuation of the previous project and built on the set of developed descriptors for elements of ML-enabled system, to validate these descriptors at larger scale and provide automation to improve consistent and timely detection of mismatches. A suite of tools was developed to (1) automate mismatch detection, and (2) demonstrate how the descriptors can be extended to support testing of ML components, particularly for production-readiness.

• ML Test and Evaluation (MLTE): System-centric, quality-attribute driven, semi-automated process and infrastructure that enables negotiation, specification, and testing of ML component requirements. Developed in collaboration with the Army AI Integration Center (AI2C). Available on GitHub.
• TEC: ML Mismatch Detection and Analysis tool leveraging the descriptors created in the previous project. Available on GitHub.
• UnitML: Tool for unit testing of ML Components. Available on GitHub.

Grace also led the following research projects related to Edge Software Systems and IoT Security:

Kalki: High-Assurance Software-Defined IoT Security

This project developed an IoT security platform that operates, with high assurance, in a resilient and trustworthy manner even in the presence of a powerful and realistic attacker who can compromise IoT devices, control nodes, and other intermediaries. The software-defined IoT security platform is composed of (i) a high-assurance control node that monitors security-relevant events and alters the "security postures" of IoT devices to enforce specific policies; (ii) trusted data nodes that execute these security postures for each IoT device using "micro-middleboxes." The Kalki IoT security platform is available on GitHub.

Authentication and Authorization of IoT Devices in Tactical Environments

This project evaluated adapted, and implemented an IETF proposal for authentication and authorization in constrained environments (ACE) such that it is resilient to high-priority threats of tactical environments (e.g., node impersonation and capture) that are currently not addressed in ACE. The SEI-ACE implementation is available on GitHub.

Tactical Cloudlets

Tactical cloudlets are forward-deployed, discoverable computing nodes that can be hosted on vehicles and other platforms to provide computation and data services at the edge, where there is limited (or no) connectivity to the cloud. The tactical cloudlet software is available as KD-Cloudlet on GitHub. Tactical cloudlets include a novel mechanism for secure key generation and exchange for establishing trust between cloudlets and mobile clients, as well as secure service migration between cloudlets.