DNN security from the perspective of ML infrastructures

PhD Qualifying Examination


Title: "DNN security from the perspective of ML infrastructures"

by

Mr. Yanzuo CHEN


Abstract:

Deep neural networks (DNNs) are paramount in advancing the capabilities across 
various domains of artificial intelligence like natural language processing, 
computer vision, and autonomous driving. In the realization and operational 
deployment of these DNN models, machine learning (ML) infrastructures, 
encompassing components such as deep learning (DL) frameworks, runtime 
environments, and computing hardware, play a crucial role. As such, their 
security is essential, and current research has shown that attackers can 
exploit vulnerabilities in these infrastructures to cause severe damage, 
including privacy leaks, corrupted model behaviors, and distorted outputs. 
Defenses have also been proposed to mitigate these threats and enhance the 
robustness of ML infrastructures and DNN applications.

This survey aims to provide a comprehensive review of existing research on the 
security of ML infrastructures and its interaction with DNN models and 
applications. We begin by introducing the modern ML infrastructures as they 
have rapidly evolved over the past decade. Then, we present the current attack 
and defense works, categorizing them from multiple aspects including their 
objectives, attacked and defended targets, and the underlying techniques. We 
conclude by discussing the limitations and extensions of existing research, 
providing insights into future research directions in this field.


Date:                   Monday, 13 May 2024

Time:                   3:00pm - 5:00pm

Venue:                  Room 3494
                        Lifts 25/26

Committee Members:      Dr. Shaui Wang (Supervisor)
                        Dr. Binhang Yuan (Chairperson)
                        Dr. Dimitris Papadopoulos
                        Dr. Wei Wang