Performance and Robustness-Aware Deep Neural Network Compression for Next-Gen AI Accelerators

Deep Learning (DL) is seen by many as “the” framework to solve complex problems involved in numerous applications that influence people’s lives. It is widely used in safety and security-critical environments such as self-driving cars as well as drones and robotics. This makes it highly essential to secure DL algorithms and systems from malicious actors. To address this, recent work has focused on developing efficient and effective techniques to defend against such attacks on deep neural networks. One major challenge that state-of-the-art defense methods commonly face is that they often study large networks, even for simple tasks. This is impractical to deploy, especially in resource-constrained settings. This project evaluates and explores different methods to decrease the storage and computational costs of robust deep neural models by exploring the impact of model compression techniques such as quantization and pruning on network performance and adversarial robustness.

Faculty Supervisor:

Anthony Bonner

Student:

Partner:

AMD Canada

Discipline:

Computer science

Sector:

Manufacturing; Professional, scientific and technical services

University:

University of Toronto

Program: