演講主題：
Hiding a Swarm's Leader from RL Agent and Human
AI Safety in the Information and Physical Spaces
講者:
Prof. Michael Lewis 美國匹茲堡大學（University of Pittsburgh）資訊學院教授
School of Computing and Information,
Department of Informatics and Networked Systems, University of Pittsburgh
Prof. Katia Sycara 美國卡內基美隆大學（CMU）機器人研究所資深研究教授，現任機器人研究所副主任，並擔任可信任人機協作研究中心主任
Edward Fredkin Research Professor of Robotics, School of Computer Science,
Carnegie Mellon University
時間: 2026年5月12日 (二) 15:00-17:00
地點: 管理大樓11樓-AI講堂
直播連結: https://reurl.cc/xW6OmZ或 掃描海報 QR code
講者簡介:
Michael Lewis 教授為美國匹茲堡大學（University of Pittsburgh）資訊學院教授，專長於人機互動、人機協作與群體機器人（swarm robotics）研究，特別關注人類如何與多機器人系統及AI協同運作。其研究涵蓋人因工程、強化學習、人機信任與決策支援等領域，長期獲美國國防高等研究計畫署（DARPA）與多個政府機構支持，為人機協作與智慧系統領域的重要學者。
Michael Lewis is a Professor at the School of Computing and Information at the University of Pittsburgh. Trained in engineering psychology, his research focuses on human-computer interaction, human-agent teaming, and swarm robotics. He investigates how humans interact with complex autonomous systems, with particular emphasis on trust, decision-making, and coordination in multi-agent environments.
His work integrates artificial intelligence, visualization, and human factors to improve the effectiveness of human–AI collaboration. Professor Lewis has led and contributed to numerous research projects supported by agencies such as DARPA, NSF, and other U.S. government organizations. He has published extensively in leading journals and conferences in human-machine systems, robotics, and AI.
Katia Sycara 教授為美國卡內基美隆大學（CMU）機器人研究所資深研究教授，現任機器人研究所副主任，並擔任可信任人機協作研究中心主任。她為人工智慧、多代理系統與人機協作領域的國際權威學者，為 AAAI 與 IEEE Fellow，並曾獲多項國際學術終身成就獎。其學術影響力深遠，發表超過700篇論文，引用數逾5萬次，致力於發展可信任、自主且可解釋的AI系統。
Katia Sycara is the Edward Fredkin Research Professor in the School of Computer Science at Carnegie Mellon University and a Research Professor at the Robotics Institute. She is a leading expert in artificial intelligence, multi-agent systems, and human-agent collaboration.
Her research spans AI autonomy, distributed intelligent systems, and trust in human-AI interaction, with applications in robotics, defense, and large-scale information systems. She has been a pioneer in agent-based systems and has made significant contributions to semantic web technologies and collaborative AI.
Professor Sycara has received numerous honors, including recognition as a Fellow of the Association for the Advancement of Artificial Intelligence and the Institute of Electrical and Electronics Engineers. She has also served in key advisory roles for government and international research initiatives, and her work has had a lasting impact on the development of intelligent, cooperative systems.
演講大綱:
Hiding a Swarm's Leader from RL Agent and Human
本演講探討在機器人群體系統中保護領導者的策略。雖然以領導者為核心的控制方式可提升群體協作效率，但同時也增加系統遭受攻擊的風險。本研究透過圖神經網路（GNN）訓練群體追隨領導者，並利用對抗式模型辨識領導者位置。研究結果顯示，在一般情境下，AI模型在領導者辨識上優於人類；然而當群體採用隱匿策略時，人類的辨識能力反而優於AI。即使在對抗模型持續學習及高視覺干擾環境下，人類仍展現出穩定且較佳的判斷能力。此結果突顯人類與人工智慧在複雜多代理系統中的關鍵差異。
This talk explores methods for protecting leadership within robotic swarms, where leader-based control improves coordination but introduces vulnerability. Using graph neural networks (GNNs), swarms can be trained to follow a leader, while adversarial models attempt to identify it. Results show that although AI models outperform humans in identifying leaders under normal conditions, humans become more effective when swarms adopt deception strategies to hide leaders. Even with adaptive adversaries and increased visual complexity, human observers demonstrate robust performance. These findings highlight key differences between human perception and AI in complex multi-agent environments.
AI Safety in the Information and Physical Spaces
本演講探討基礎多模態模型（Foundational Multi-Modal Models）的潛在安全漏洞，重點分析透過「意圖欺騙（intention deception）」進行模型越獄（jailbreaking）的風險，以及在何種條件下多模態模型更容易受到對抗性攻擊並洩露敏感或危險資訊。此外，將提出相應的安全防護策略，並進一步介紹「情境安全（contextual safety）」在真實世界應用中所面臨的挑戰。
In this talk, we will present our work on vulnerabilities of Foundational Multi-Modal Models. In particular, we will present jailbreaking of Frontier Models via intention deception and conditions that make multi modal models more vulnerable to adversarial attacks in disclosing dangerous information. We will also propose safety mitigations. Additionally, we will introduce contextual safety and its challenges in the physical world.
※本活動無需報名
