Title:
Enabling adversarial detection and robust inference for energy-efficient and secure edge AI
Abstract:
Fine-tuning pretrained models has become a widespread solution to deploy vision AI systems. The main reason that makes this approach popular is the fact that it provides an efficient, low-cost avenue to adapt generic convolutional models to various vision tasks from image recognition, to object detection and semantic segmentation. Unfortunately, securing a model against adversarial attacks has not followed the same trend, and model strengthening remains expensive and specialized.
In this talk I will first introduce an energy-efficient solution for multi-task inference at the edge. I will then present a novel approach that applies a similar task adaptation technique to adversarial sample detection and robustness. Finally, I will show some hardware optimizations that can be applied to efficiently deploy these models on resource-constrained devices for edge AI.
Speaker:
Marco Donato
Assistant Professor, Department of Electrical and Computer Engineering, Tufts University
Bio:
Marco Donato is an Assistant Professor in the Department of Electrical and Computer Engineering with a secondary appointment in Computer Science at Tufts University. Prior to joining Tufts, he was a Postdoctoral Fellow in the John A. Paulson School of Engineering and Applied Sciences at Harvard University. He earned a Ph.D. in Electrical Sciences and Computer Engineering from Brown University and holds a M.Sc. and B.Sc. degree from the University of La Sapienza in Rome, Italy.
Dr. Donato is the head of the Tufts Emerging Circuits and Systems (TECS) Lab, where he leads a team working at the intersection of VLSI circuits and computer architecture to design energy-efficient, reliable, and secure computing systems. Their research focuses on developing cross-stack co-design methodologies, with particular emphasis on heterogeneous memory integration and non-volatile memories. Their goal is to enable the integration of emerging technologies to address the challenges of technology scaling and growing application complexity in modern computing systems.
Host: Professor John McNeill