No One Machine Learning Model to Rule Them All

ABSTRACT

With the rapid rise of deep learning and generative AI, there is a strong push to deploy increasingly complex machine learning models for critical tasks in cyber and tactical edge domains. However, in resource-constrained environments, where both data and compute are limited, traditional one-ML-model-fits-all approaches fall short. In this talk, we argue for a dynamic, adaptive framework in which multiple models are trained, selectively deployed based on available resources, and continuously updated to withstand adversarial attacks and respond to rapidly changing environments. We will present our ongoing work toward realizing this vision and demonstrate how such an approach can enable resilient, efficient, and mission-ready AI at the edge, while also examining the role of human involvement in enhancing system robustness and reliability.

BIOGRAPHY

Murat Kantarcioglu is a Professor of Computer Science and a CCI Faculty Fellow at Virginia Tech. He earned his PhD in Computer Science from Purdue University, where he received the Purdue CERIAS Diamond Award for Academic Excellence. He is currently affiliated with Harvard University’s Data Privacy Lab as a faculty associate and previously served as a Visiting Scholar at UC Berkeley’s RISELab. Dr. Kantarcioglu’s research focuses on integrating cybersecurity, data science, and blockchain technologies to develop secure and efficient mechanisms for data processing and sharing. His work has been featured by media outlets including the Boston Globe, ABC News, PBS/KERA, Associated Press, BBC, and the Guardian, and he has received multiple best paper awards. His honors include the NSF CAREER Award, the AMIA 2014 Homer R. Warner Award, and the IEEE ISI 2017 Technical Achievement Award for contributions to data security and privacy. He is a Fellow of the AAAS, ACMI, and IEEE.