Title:
Trustworthy Distributed Learning in Uncertain and Adversarial Environments
Abstract:
Distributed learning enables large populations of devices and agents to collaboratively train machine learning models without centralizing data. However, real-world deployments must contend with unreliable communication, intermittent participation, system heterogeneity, privacy constraints, and adversarial attacks. These challenges can significantly degrade the reliability, efficiency, and security of existing learning algorithms.
In this talk, I will present recent advances in designing provably robust distributed learning algorithms under such uncertainties. Specifically, I will discuss methods for handling dynamic communication failures, arbitrary participant unavailability, and the joint challenges of privacy preservation and adversarial robustness. Together, these results provide theoretical foundations and practical algorithms for trustworthy distributed learning in uncertain and adversarial environments.
Speaker:
Lili (Liliana) Su
Assistant Professor, Northeastern University
Bio:
Lili (Liliana) Su is an Assistant Professor at Northeastern University. Her research lies at the intersection of machine learning, distributed systems, reinforcement learning, and autonomous systems, with a focus on developing trustworthy AI that can learn, adapt, and operate reliably under uncertainty, heterogeneity, and adversarial conditions. Her work has been supported by NSF, ONR, ARL, ARPA-H, and industry sponsors.
Host: Professor John McNeill