:::

TIGP (AIoT) Seminar -- With the advance of text-to-image generative models nowadays, we will discuss the security mechanisms and vulnerabilities related to these models, followed by introducing a red-teaming approach to automatically discover the problematic text prompts for attack the security mechanisms. Additionally, we will introduce how to transform vision-language models into deepfake image detection tools which are more general for various types of deepfake images


  • 講者 : 邱維辰 教授
  • 日期 : 2024/10/04 (Fri.) 14:00~16:00
  • 地點 : 資訊所新館106演講廳
  • 邀請人 : TIGP (AIoT)
Abstract
This talk provides a comprehensive exploration of mixed reality (MR), delving into its evolution, challenges, and future prospects. We go through the developmental history of MR goggles, tracing their transformative advancements and identifying key hurdles yet to be overcome. While remote rendering presents a promising solution for lightweight MR goggles, latency issues persist. By leveraging server object streaming and real-time pose assistance, we aim to mitigate these concerns, offering users an outstanding experience. Additionally, network traffic optimization techniques are proposed to enhance system scalability. This architectural framework not only achieves low latency but also supports various terminal devices in multi-user, real-time, and MR scenarios. Looking forward, we outline potential research directions, envisioning the continued evolution of MR technology and its integration into broader applications.