From Pixels to Agents: Building Trustworthy and Secure AI
演讲者
时间
2026年11月05日 15:00 至 16:00
地点
A3-2-303
线上
Zoom 242 742 6089
(BIMSA)
摘要
As AI becomes part of scientific research and everyday decision-making, strong performance alone is not enough: systems must also resist manipulation and remain reliable when deployed. This talk introduces research on trustworthy and secure AI through three connected questions: How can AI be misled? How can it be protected? And how can users check that a deployed system has not been altered? Accessible examples spanning image recognition, conversational assistants, and AI systems that search for information and carry out tasks illustrate both emerging risks and practical approaches to addressing them. The aim is to connect these challenges with problems in other disciplines and explore opportunities for collaboration on realistic applications, shared evaluation methods, and the mathematical foundations of trustworthy AI.
演讲者介绍
朱斌于2001年12月至2026年6月在微软工作,其中约22年在微软亚洲研究院,之后在M365 Core和Bing Core Search产品团队任职。他曾在微软亚洲研究院及产品团队担任首席研究员。他的研究涵盖可信AI、AI安全、AI在安全中的应用、可信计算、隐私保护计算、用户认证、验证码、Web系统和多媒体安全。近期工作包括AI安全与隐私保护研究,尤其是LLM与智能体安全、模型篡改检测与知识产权保护、对抗鲁棒性和隐私保护计算;同时也包括AI助手推理、深度搜索、评估、提示词优化、文档理解和多向量检索等方向。其研究成果已转化到微软产品和服务,包括人工操作勒索攻击防御、恶意PowerShell脚本检测、钓鱼与欺诈检测、CAPTCHA、Bing索引压缩和移动完整性检测等。他已发表125余篇论文,拥有58项已授权美国专利,并长期参与学生培养及与中国、美国和英国高校的研究合作。