Lopata Hall, Room 101
"Targeted GAN: Robust Classification with Targeted Adversarial Learning"
Adviser: Yixin Chen
Deep neural networks (DNNs) are highly expressive models while at the same time, are susceptible to widely existed adversarial examples: such as images intentionally modified with imperceivable changes to deceive machine learning models. In addition, adversarial examples have transferability, being able to fool across models of totally different architectures. How to generate high quality adversarial examples and how to increase models' resistance to adversarial examples become two main directions. In this paper, we propose a novel training paradigm called targeted GAN (tGAN) that can train a generator G with the ability to produce adversarial examples with target labels, and simultaneously cultivate a state-of-the-art classifier C that is more robust to imperceivable adversarial examples. G keeps track of adversarial information provided by a differentiable classier C through back-propagation and uses this information to further train the classifier C. Extensive experimental results demonstrate two distinctive benefits of tGAN. First, our framework can be applied to any existing deep learning models and helps increase their robustness to adversarial examples and reduce their test error at the same time. Second, the generator G in tGAN can synthesize human-indiscernible adversarial examples with only a single forward pass for any target label, providing high-quality adversarial attacks which can be used to improve the robustness of other DNNs.
"RTM: a Scalable Real-Time Messaging Middleware"
Adviser: Chenyang Lu
Cloud platforms must support numerous distributed (micro-) services with diverse Quality of Service (QoS) requirements. Cloud-based services often communicate with each other through a messaging middleware based on the publish/subscribe paradigm. Messaging middleware has become a critical Platform as a Service (PaaS). Recently, with the emergence of Industry 4.0 applications and 5G core networks, it becomes critical to support real-time messaging services between latency-critical services in cloud environments. A messaging middleware must meet diverse QoS (latency/traffic rate) requirements in face of a large number of concurrent connections. To tackle these challenges, we are developing Real-Time Messaging (RTM), a scalable real-time messaging middleware for cloud applications. At the message level, RTM supports isolated, differentiated service and distributed deployment. This talk will present the design and preliminary evaluation of RTM for (1) latency differentiation, (2) performance isolation and 3) scalability to concurrent connections through load distribution and distributed rate control.