Machine Learning for Secure and Resilient Information Management in Combat Cloud

Sole PI: Sanjay Madria

Amount: $500,000

Time period: 1/2021 - 05/2022

In a battlefield zone, Command-and-Control (C2) capabilities can be improved using the Combat Cloud paradigm which can enable network-aware disruption tolerant information flows, provide distributed data stores, and efficient data dissemination across several groups of forces deployed with different mission goals. To continue battlefield missions appropriately, and get a better understanding of the situation, the forces as well as C2 need to collect timely information generated in the war zone. However, due to the damaged/degraded network infrastructure, or the unavailability of information-servers connectivity especially in the hostile area, it is a challenge to forward information in this extreme situation. In this dynamic surrounding, any sudden important event-related information or possible mission updates (as the mission evolves) should also be sent to C2 with the help of intermediate nodes regardless of their mission interests. We propose to learn the mission interests dynamically, and optimally store and forward the information generated by the nodes as the mission evolves to C2 using Reinforcement Learning (RL). In this forwarding process, we will focus on identifying the trending mission interests (related to updated missions or events) for continuous decision-making by considering the mobility and connectivity of the nodes and considering the changes of the perceived network model and accordingly modify appropriate ‘reward’ functions based on past learning. The machine learning will help in learn-and-adapt to space-time evolution of data requests as well as local policies used by at nodes in determining the currently cached objects and what should be prefetched next along with determining how many objects need to be prefetched based on determined mission priorities, expected latency, etc. These features, in practice, usually exhibit unknown and temporal dynamics because the most popular content at the current epoch may not receive the highest attention in the future; and mobile users could change locations as time passes. The combat cloud in contested environments faces challenges while making the prioritized data available to different groups in a timely and secure fashion (authenticated, un-tempered, and trusted). For secure, end-to-end, mission-oriented, data dissemination, it needs a resilient and secure information processing layer for Information Exchange Requirements (IERs) (tasks, operational elements, and information flow). The security and resiliency need machine learning-based methods for targeted content dissemination, and, proactive dissemination/caching (TA1), and dynamic mission-oriented data discovery (TA2). Secure information processing in combat cloud needs efficient and dynamic fine-grained Attribute-based key distribution, verification, and revocation for group-based coordination (TA3) for a collaborative DIL environment. We will design algorithms, and develop a system prototype in a Delayed/Disconnected, Intermittently-Connected, Low-Bandwidth (DIL) environment to validate the Combat Cloud design discussed here.

CRII: OAC: Enabling Quantities-of-Interest Error Control for Trust-Driven Lossy Compression

PI: Xin Liang

Amount: $175,000

Time period: 06/01/2022 - 05/31/2024

Scientific simulations and instruments are producing data at volumes and velocities that overwhelm network and storage systems. Although error-controlled lossy compressors have been employed to mitigate these data issues, many scientists still feel reluctant to adopt them because these compressors provide no guarantee on the accuracy of downstream analysis results derived from raw data. This project aims to fill this gap by developing a trust-driven lossy data compression infrastructure capable of strictly controlling the errors in downstream analysis theoretically and practically to facilitate the use of data reduction in scientific applications. Success of this project will promote the progress of science in multiple disciplines via effective data reduction, and contribute to resolving important societal problems including electric generation, weather forecasting, material design, and transportation. Moreover, this project will contribute to the growth and development of future generations of scientists and engineers through educational and engagement activities, including development of new curriculum and recruitment of K-12 students.

Existing lossy compression techniques either overlook error quantification or provide error control only for raw data, leaving uncertainties in the outcome of downstream quantities of interest (QoIs) computed from the raw data. This greatly concerns many computational scientists who wish to reduce their data while preserving necessary information, preventing them from adopting lossy compression in their applications. This research will address these problems through an integration of theory and implementation via three tasks. First, a novel theory enabling error control on downstream QoIs will be developed. This will fundamentally address the trust issues of existing error controlled lossy compressors that provide error control only on raw data. Second, an optimization method ensuring tight error control will be applied based on rigorous analysis, to achieve higher compression ratios under the same requirements. Third, a scalable infrastructure will be built through a careful integration with advanced compression frameworks and tailored parallelization based on target QoIs, in order to take full advantage of existing compression algorithms and computational patterns in the target QoIs. The project will enable application scientists to store the most valuable information in their data based on their unique needs, creating opportunities for novel findings in multiple scientific disciplines including climatology, cosmology, and seismology.