Robust Processing Techniques for Complex RF Applications Using Generative AI/ML Techniques in the Presence of Training-Testing Distribution Mismatch

OUSD (R&E) CRITICAL TECHNOLOGY AREA(S): Trusted AI and Autonomy The technology within this topic is restricted under the International Traffic in Arms Regulation (ITAR), 22 CFR Parts 120-130, which controls the export and import of defense-related material and services, including export of sensitive technical data, or the Export Administration Regulation (EAR), 15 CFR Parts 730-774, which controls dual use items. Offerors must disclose any proposed use of foreign nationals (FNs), their country(ies) of origin, the type of visa or work permit possessed, and the statement of work (SOW) tasks intended for accomplishment by the FN(s) in accordance with the Announcement. Offerors are advised foreign nationals proposed to perform on this topic may be restricted due to the technical data under US Export Control Laws. OBJECTIVE: Develop efficient techniques to support Generative AI/ML algorithms for RF applications in the presence of severe mismatches in the distributions of the data between training and live operations. These techniques should rapidly detect anomalies in the data and provide corrective measures to assist the Generative AI enhanced AI/ML algorithms to regain the lost performance. The robustness of these techniques to various operating conditions should be demonstrated using a high-fidelity RF Digital Twin for an advanced RF application of relevance to the Air Force. DESCRIPTION: Training data mismatches have been studied in the RF domain for techniques such as STAP [1]. The increasing use of AI/ML techniques for RF applications that are typically operating in data-starved environments has further highlighted the importance of Digital Engineering tools that can capture and model the real-world effects in a synthetic environment. While state-of-the-art Digital Twins [2] faithfully recreate most of the realistic physics-based phenomenon, it is inevitable that unmodeled or unknown characteristics can end up in the data when the algorithms are exposed to live operations. It is important to develop computationally tractable measures that can detect these anomalies in the data and apply corrections and/or identify severely-out-of-training-distribution scenarios to ensure the AI/ML is not derailed by these unmodeled mismatch between training and testing data. While classical textbook methods for anomaly detection and classification are computationally scalable, they rely heavily on Gaussianity assumptions which limit their applicability to the envisioned heterogeneous and cluttered operational scenarios. Bayesian and Monte-Carlo techniques suffer from the curse of high dimensionality and are sensitive to parameter selection in their training (e.g. kernel width selection). Generative AI techniques, including GANs and Diffusion Models [3.4] have recently demonstrated the ability to model complex high dimensional distributions using sufficiently rich training data in ways that can be used to generate new data samples. Their utility in the context of testing for or identifying distributional shifts remains to be fully established. It is important that the computational methods proposed are scalable to high dimensional settings and are able to quantitatively know what they don't know . [5] The main deliverables on this project will be advanced techniques that leverage or build on recent advances in generative AI to detect and classify anomalies and distributional shifts in the live/testing data compared to the training datasets, corrective algorithms, and demonstration of this novel approach on an advanced RF application using data from an RF Digital Twin as well as theoretical and/or computational analysis of their associated fundamental inferential limits. PHASE I: This topic is intended for technology proven ready to move directly into Phase II. Therefore, a Phase I award is not required. The offeror is required to provide detail and documentation in the Direct to Phase II proposal which demonstrates accomplishment of a Phase I-like effort, including a feasibility study. This includes determining, insofar as possible, the scientific and technical merit and feasibility of ideas appearing to have commercial potential. It must have validated the product-market fit between the proposed solution and a potential AF stakeholder. The offeror should have defined a clear, immediately actionable plan with the proposed solution and the AF customer. Relevant areas of demonstrated experience and success include high-fidelity M&S, solutions to complex RF problems using AI/ML, concept development, concept demonstration and concept evaluation. PHASE II: As this is a Direct-to-Phase-II (D2P2) topic, no Phase I awards will be made as a result of this topic. To qualify for this D2P2 topic, the Air Force expects the applicant(s) to demonstrate feasibility by means of a prior Phase I-type effort that does not constitute work undertaken as part of a prior or ongoing SBIR/STTR funding agreement. These efforts will include developing a high-fidelity physics-based M&S, simulation of an AI/ML relevant solution for an AF application using the M&S tool, demonstration of the vulnerability of these techniques to model mismatches, and a practical analysis on computationally tractable techniques to overcome these mismatches. PHASE III DUAL USE APPLICATIONS: The proposer will identify potential commercial and dual use applications for this technology. REFERENCES: 1. M. Rangaswamy, B. Himed, and J.H. Michels, Statistical analysis of the nonhomogeneity detector for STAP applications, Digital Signal Processing, vol. 14, no. 3, May 2004, pp. 253-267; 2. S. Gogineni, J. R. Guerci, H. K. Nguyen, J. S. Bergin, D. R. Kirk, B. C. Watson, and M. Rangaswamy, High fidelity RF clutter modeling and simulation, IEEE Aerospace and Electronic Systems Magazine, Vol. 37, pp. 24-43, Nov 2022; 3. San-Roman, Robin, Eliya Nachmani, and Lior Wolf. "Noise estimation for generative diffusion models." arXiv preprint arXiv:2104.02600 (2021); 4. Song, Yang, et al. "Score-based generative modeling through stochastic differential equations." arXiv preprint arXiv:2011.13456 (2020); 5. Nalisnick, Eric, et al. "Do deep generative models know what they don't know?." arXiv preprint arXiv:1810.09136 (2018). KEYWORDS: Robust Processing; Generative AI/ML; Outlier detection