千善日-面向多领域的AI稳健性评估技术与案例分析.pdf

1、面向多领域的AI稳健性评估技术与案例分析千善日(Chon sun il)|ThinkforBL千善日(Chon Sun il)Manager of ThinkforBL Co.,Inc.Lead author of the AI Trustworthiness Development Guide,published by the Korean Ministry of Science and ICT,covering all domains:Smart Policing,Hiring,Generative AI,Autonomous Driving,Healthcare,and Public&S

2、ocial Services-Compatible with 63 out of 67 detailed verification items in the AI RMF published by the U.S.NISTDeveloped AI Trustworthiness Verification Techniques and established seven group standards with Koreas Telecommunications Technology Association(TTA)Certified Auditor for AI Management Syst

3、em(ISO/IEC 42001)Certified in Functional Safety Verification Frameworks AFSP(Automotive Functional Safety Professional),CACSP(Certified Automotive Cyber Security Professional)Masters Degree in Electronic Engineering from JeonbukNational University目 录CONTENTSI.AI稳健性（Robustness）评估的工程问题定义II.AI稳健性评估的技术方

4、法III.韩国的AI稳健性评估技术与标准化案例IV.真实试点项目与公共数据诊断案例V.国际扩展性与合作方向The colors of the bounding box keep on changingfor the same object,which is a detection error同一个物体的边框颜色不断变化，说明存在检测错误Performance on the right side is much better than on the left side右侧模型的识别效果明显优于左侧Load TruckTrailer TruckBusMini TruckCarThink for a

5、 Better Life Public institution shared 50,000 images for training vehicle detection 某公共机构提供了 50,000 张图像,用于车辆类型识别的训练Used50,000images|使用了 50,000张图像Used only 2,000images from the 50,000|只用了50,000张图片中的2,000张面向多领域的AI稳健性评估技术与案例分析 PART1 开开发完成Redundant datacan actually increase bias|重复数据反而可能加剧偏见 Balanced Data|均衡数据Think for a Better Life面向多领域的AI稳健性评估技术与案例分析 PART1 开开发完成扫码领取会议PPT资料感谢聆听！