Multi-Scale Hydra- A Dynamic Hybrid Submodular Framework for Multi-Scale Network Resilience Incorporating Invariant Properties
Abstract
Complex networks are pervasive in critical infrastructures such as energy, transportation, and cyber systems. In order to operate successfully in uncertain and adversarial environments, complex networks must mitigate natural and malicious disruptions. The goal of network resilience is to develop mitigation strategies that ensure performance, connectivity, and functionality during and after a disturbance. Such mitigation strategies will need to incorporate both discrete (locations of nodes-edges) and continuous (control inputs) decision variables that affect network dynamics at multiple time-scales, creating a hybrid system. We propose to investigate and develop a dynamic hybrid submodular framework for network resilience across multiple time-scales. We will identify fundamental properties (e.g., safety, stability, connectivity) that need to be be maintained under topology changes, failures, and attacks, which we term resilient network invariants. To ensure computational tractability, we will explore hybrid submodular structures of the invariants, which generalize submodular optimization techniques from discrete functions to functions with both discrete and continuous variables. We will develop submodular techniques to model and mitigate large statistical models such as Deep Neural Networks that are used for perception and decision-making by complex networks. Our research thrusts are summarized as follows- Resilient Network Invariants- We will investigate and formulate invariant properties of complex networks, as well as develop efficient algorithms for computing the invariants. We will address the research questions- (i) How to formulate invariant properties of node states, such as safety, stability, and performance. (ii) How to characterize invariant properties of the network topology, such as connectivity and capacity. (iii) How to develop distributed and scalable algorithms for computing network invariants. and (iv) How to model the impact of attacks and mitigation mechanisms on the invariants. Dynamic Hybrid Submodularity for Resilience- We will investigate and develop hybrid submodular algorithms for computing mitigation strategies that ensure network resilience. We will address the research questions- (i) How to characterize hybrid submodular properties of network invariants. (ii) How to develop minimum-cost mitigation strategies that guarantee network invariants by exploiting hybrid submodularity. (iii) How to develop distributed hybrid submodular algorithms for dynamic networks. and (iv) How to enhance scalability and ensure optimality under partial network observability. Multi-Time-Scale Resilient Decision-Making- We will develop hybrid submodular algorithms for resilient decision-making at multiple time-scales. We will address the research questions- (i) How to model network invariants across multiple time-scales. (ii) How to formulate multi-scale hybrid submodularity properties. (iii) How to develop hybrid submodular algorithms for computing multi-scale mitigation strategies. and (iv) How to develop optimal mitigation strategies against malicious attacks. Submodularity and AI Trojan Detection- We will develop a submodular optimization approach to detecting AI Trojans. We will address the research questions- (i) How to characterize submodularity of the attack effectiveness in terms of the fraction of embedded Trojan data samples. (ii) How to model the impact of AI Trojans on DNN outputs and network invariants. (iii) How to ensure resilient network operation in the presence of AI Trojans. and (iv) How to mitigate Trojan attacks via distributed learning. Testing and Validation- We will validate our algorithms on multiple datasets, such as text, image, power, and UAVs, and transition to DoD labs in consultation with the PM.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 29, 2024
- Source ID
- fa95502310208
Entities
People
- Radha Poovendran
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Washington