ECE Distinguished Seminar
Robust and Efficient Deep Learning Systems at Edge
Dr. Xue (Shelley) Lin
Department of Electrical and Computer Engineering
Northeastern University
February 25, 2022, 11:00 am
RSCH 163
Zoom Link: https://gmu.zoom.us/j/93244528991
Note
for people joining in-person only:
Abstract
This talk presents recent work from Dr. Lin’s group on deep learning security and hardware acceleration. The first part is about vulnerability of deep neural networks. The second part is to implement efficient
deep learning systems at edge covering both inference and training.
The talk begins with our design of structured adversarial examples, revealing structural information through strong group sparsity and providing better interpretability of the adversarial examples. Next, I will
introduce our adversarial T-shirt, the first physical world adversarial example considering deformation of non-rigid objects. This work was published as a Spotlight Paper in ECCV’20 and has been broadly featured and cited in over 100 media outlets including
Communications of the ACM, The Register, Boston Globe, etc. Then, I will present our recent work in ICLR’22 on reverse-engineering of adversarial perturbations to recover the original images. From hardware perspective, deep learning systems are subject to
fault injection attacks, which manipulate neural network models for misclassification. I will discuss our modeling of such attacks through ADMM (alternating direction method of multipliers).
For efficient implementation techniques of deep learning at edge, I will briefly go through our paper leveraging model compression for mobile acceleration of 3D convolutional neural networks, achieving real-time
execution for the first time. Then I will introduce our series of work on deep learning quantization for FPGA using intra-layer mixed schemes and multiple precisions. The talk ends with our NeurIPS’21 Spotlight Paper about sparse training using mobile devices.