Deep Learning Model Backdoor Detection and Removal

Speaker: Professor Xiangyu Zhang
         Samuel Conte Professor
         Department of Computer Science
         Purdue University

Title:  "Deep Learning Model Backdoor Detection and Removal"

Date:   Wednesday, 13 December 2023

Time:   4:00 pm - 5:00 pm

Venue:  Room 1409 (near lift 25/26), HKUST

Abstract:

A backdoor attack aims to induce model misclassification by stamping a
specific pattern to an input. Such a pattern can cause a large number of
inputs (of a victim class) to be misclassified to a target class Backdoors
could be injected by data poisoning during training, neuron hijacking in
existing models, or even naturally exist in pre-trained models. This talk
will discuss how to detect and remove backdoors in different
domain-specific models: computer vision, object detection, large language
models, and code-language models, driven by the experience in various AI
backdoor scanning competitions.


*******************
Biography:

Professor Xiangyu Zhang is a Samuel Conte Professor at the Computer
Science Department of Purdue University. His research focuses on program
analysis, deep learning security, and software engineering. He has
received a number of prestigious awards, such as ACM SIGPLAN Distinguished
Dissertation Award, ACM SIGPLAN Distinguished Paper Award, ACM SIGSOFT
Distinguished Paper Award, ACM CCS Best Paper Award, USENIX Security Best
Student Paper Award, NDSS Distinguished Paper Award, and ASE Best Paper
Award. He also co-supervised a Ph.D. dissertation that received the 2017
ACM SIGACT Distinguished Dissertation Award.