Collaborative Research: SaTC: CORE: Small: Corruption-Robust Online Optimization: from Theory to Applications

University Of Massachusetts Amherst, Amherst MA

Investigators

Abstract

From online advertising to content delivery networks and recommendation systems, many modern technologies rely on algorithms that must operate in real time without knowing what will happen next. Researchers in computer science, operations research, engineering, and other fields have developed powerful online optimization and learning tools for making effective decisions in the face of uncertainty, by helping systems learn from past outcomes to improve performance over time. However, these methods are vulnerable to attackers aiming to disrupt the system: fake reviews, ad fraud, denial-of-service attacks, and other attacks can corrupt the algorithms' learning process. Researchers have begun developing "corruption-robust" learning algorithms that are more resilient to attacks; however, significant barriers remain in translating these theoretical advances into real-world systems. This project aims to reduce those barriers by designing learning algorithms that are both theoretically sound and practical to implement, enabling more robust decision-making in real-world applications. This work directly supports the national interest by strengthening the resilience of critical cyberinfrastructure and advancing the scientific foundations of trustworthy AI. This project advances the theoretical foundations of online optimization and learning by explicitly incorporating robustness to security threats and adversarial corruptions. While online learning has been extensively studied for decades across computer science, control theory, economics, and operations research, a rigorous understanding of its vulnerabilities and resilience under malicious attacks remains underdeveloped. This project fills this gap by developing new corruption-robust algorithms for sequential decision-making problems, particularly in settings involving combinatorial action spaces and resource constraints in both single-agent and multi-agent learning environments, addressing various classes of adversarial threats. The research involves designing algorithms with provable performance guarantees in the presence of corruption, analyzing their theoretical properties, and developing efficient implementations. To validate its impact, the project evaluates the proposed methods in two real-world applications: probabilistic maximum coverage for content delivery networks and online learning to rank in social and e-commerce platforms. The project's outcomes will contribute broadly to secure and trustworthy decision-making systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.