演講主題:可信賴人工智慧與數位媒體鑑識: 應對模型的自我感知,真實性與安全挑戰
講者: 張明清 紐約州立大學奧本尼分校電腦科學系 副教授
時間: 2024/11/05 (二) 14:10-16:00
地點: 管理大樓11樓-AI講堂
直播連結: https://gqr.sh/LrGY
講者簡介:
Ming-Ching Chang is an Associate Professor with tenure (since Fall 2022) in the Department of Computer Science at the University at Albany, SUNY. He previously held positions in the Department of Electrical and Computer Engineering (2016-2018) and as an Adjunct Professor in Computer Science (2012-2016). From 2008 to 2016, he worked as a Computer Scientist at GE Global Research Center and was an Assistant Researcher at the Mechanical Industry Research Labs, ITRI in Taiwan from 1996 to 1998.
Dr. Chang earned his Ph.D. in Engineering Man/Machine Systems from Brown University in 2008, along with an M.S. in Computer Science and Information Engineering (1998) and a B.S. in Civil Engineering (1996) from National Taiwan University. His research focuses on video analytics, computer vision, image processing, and artificial intelligence, with over 70 published papers. His projects have received funding from DARPA, IARPA, NIJ, VA, GE Global Research, Kitware Inc., and the University at Albany. Dr. Chang is a senior member of IEEE.
演講大綱:
可信賴人工智慧的研究目標是創建高效、健全、安全、公平、隱私保護和可問責的人工智慧模型。 隨著基礎模型和生成式 AI 的應用日益增長,這些技術可以生成文件與產生可以以假亂實的圖像。在快速變化的數位世界中,真實與虛假的界線變得愈發模糊。對於能夠驗證媒體內容、辨別真偽的技術需求日益重要。
在此次演講中,我將探討可信賴人工智慧、數位媒體鑑識以及安全計算領域的創新研究成果。首先,我將介紹一種新穎的多流形嵌入學習方法,用於識別分佈外數據(OOD),並通過因果分析揭示大型視覺語言模型中的隱藏幻覺因素。此外,我將討論一種針對長尾數據分佈的噪聲標籤學習技術。
在數位媒體鑑識領域,我將展示使用隱式神經表徵進行圖像操作檢測 (IMD),這些技術在有限監督下進行。此外,我會闡述 IMD 數據集的開發, 包括物體感知和語義重要性的標註,並通過穩定擴散 (Stable Diffusion) 技術更有效地模擬現實場景。
最後,我將探討安全加密計算中的重要創新,特別是利用GPU加速深度神經網路推理中的完全同態加密 (FHE),以及通過量化和網路微調策略來增強功能引導。
主辦單位:智慧運算學院、人工智慧研究中心
※本活動無需報名