Seminar: Exploring the DNN Memorization Effect in Transparent Computing

Abstract: Modern computing systems are complex and opaque, posing significant security risks. Transparent computing aims to expose these black boxes to better comprehend system behaviors. Recent advances in artificial intelligence (AI) have altered the functioning of modern computing systems, creating new challenges and opportunities for transparent computing. On the one hand, many AI systems are black boxes with dense connections among their computing units, rendering existing techniques like dependency analysis inadequate. This highlights the need for novel methods to improve transparency and defend against deep neural network (DNN) attacks. On the other hand, AI provides a new computation abstraction that facilitates data-driven computation-heavy applications. This opens up new possibilities for transparent computing, particularly in large-scale data processing systems. In this talk, I will present my work in these two directions, concentrating on the DNN memorization effect. First, I will discuss how DNN memorization leads to Trojan attacks and introduces new challenges in analyzing deep neural networks for security purposes. I will introduce our innovative approach to detecting and eliminating DNN Trojan behaviors through our analysis. Afterward, I will explain how to harness the DNN memorization effect to create efficient data storage systems for auditing data.