活动邀请 | 广东省大数据计算基础理论与方法重点实验室项目启动会暨于全院士大师讲堂

广东省大数据计算基础理论与方法重点实验室项目启动仪式暨大师讲堂“蜂群民主决策与类脑计算”将于7月20日下午14:00-16:45在行政楼W201举行。

当天下午14:30-15:30，我国无线通信领域的学科带头人于全院士将做客香港中文大学（深圳）“大师讲堂”。他将以“蜂群民主决策与类脑计算”为主题开展报告，与我们分享科研的心得体会。

“青年讲座”将于15:45-16:45举行。我校理工学院助理教授韩晓光和数据科学学院副教授吴保元将分别带来“基于显式和隐式表达结合的可视媒体理解”和“现实场景下的人工智能安全问题研究”主题讲座。

本次活动无需报名，我们诚邀各位老师及同学出席。

蜂群民主决策

类脑计算

活动安排

日期：2021年7月20日，星期二

时间：14:00-16:45

地点：行政楼西翼W201会议室

语言：中文

14:00-14:30广东省大数据计算基础理论与方法重点实验室项目启动仪式

14:30-15:30大师讲堂：蜂群民主决策与类脑计算

主讲人：于全院士

15:30-15:45 茶歇

15:45-16:15青年讲座1

主题：基于显式和隐式表达结合的可视媒体理解

主讲人：韩晓光教授

16:15-16:45 青年讲座2

主题：现实场景下的人工智能安全问题研究主讲人：吴保元教授

Event Rundown

Date: Tue. July 20, 2021

Time:2:00 pm – 4:45 pm

Venue:W201, Administration Building

Language:Chinese

14:00-14:30 Guangdong Provincial Key Laboratory of Big Data Computing Launching Ceremony

14:30-15:30 Master Forum

Speaker: Academician Quan YU

Topic: Honeybee democratic decision makingandbrain-likecomputing

15:30-15:45 Tea Break

15:45-16:15 Key Lab Young Scholars Seminar 1

Speaker: Professor Xiaoguang HAN

Topic: Combing Explicit and Implicit Representation for Image, Video and 3D Understanding

16:15-16:45 Key Lab Young Scholars Seminar 2

Speaker: Professor Baoyuan WU

Topic: Research on the Security Problem of Artificial Intelligence in Real Scenarios

嘉宾简介

Speaker Profile

大师讲堂嘉宾

于全院士

系统工程研究院研究员

信息系统重点实验室主任

鹏城实验室研究员

于全院士是我国无线通信领域的学科带头人。他于2009年当选中国工程院信息与电子工程学部院士。

获奖情况：

获国家科技进步一等奖1项、二等奖1项，军队科技进步一等奖5项、二等奖2项。国家发明专利授权9项，出版学术专著9部。获得全国优秀科技工作者（1997年）、“中国科协求是杰出青年奖”（1999年）、“第九届中国青年科技奖”（2006年）、“中国青年五四奖章标兵”（2008年）、何梁何利基金“科学与技术进步奖”（2010年）以及中国政府出版奖（2018年）等荣誉称号。

研究领域：

无线自组织网络、空间信息网络、软件无线电、认知无线网、网络体系架构等。

于全院士近年来的学术兴趣主要在类脑神经元的无线异构网络和类生物免疫的网络安全防御等交叉学科上。

Swipe up to view

Academician Quan Yu has been an Academician of China Academy of Engineering (CAE) since 2009. His research area includes wireless communications and network architecture design.

Awards:

Prof. Yu received the National Science and Technology Progress Award (1st class once, 2nd class twice). He also holds 9 National Invention Patents and published 9 academic monographs. He is the recipient of a number of honors and awards: The National Excellent Technologist (1997), China Association of Science and Technology Outstanding Young Scholars Award (1999), The 9th Science and Technology Award for Youth of China (2006), China “May 4th” Prize for Young Scholars (2008), Ho Leung Ho Lee Foundation Science and Technology Progress Award (2010) and Chinese Government Press Prize (2018), etc.

ResearchAreas:

wireless ad-hoc networks, space information networks, software defined radio, cognitive radio networks and network architecture, etc.

His most recent research focuses on the interdisciplinary subjects, such as brain-inspired wireless heterogeneous networks and bio-inspired cyber defense systems, etc.

青年讲座嘉宾

韩晓光助理教授

理工学院助理教授

韩晓光博士是香港中文大学（深圳）理工学院助理教授、校长青年学者。2009年本科于南京航空航天大学毕业，2011年获得浙江大学应用数学系硕士学位，2011年至2013年于香港城市大学创意媒体学院任研究助理，之后于2017年获得香港大学计算机科学专业博士学位。他在该方向上已有40余篇论文发表于著名国际期刊和会议，包括顶级会议和期刊SIGGRAPH、CVPR、ICCV、ECCV、AAAI、ACM TOG、IEEE TVCG、IEEETIP、TPAMI等。他的工作曾获得CCF图形开源数据集奖，计算机图形学顶级会议Siggraph Asia 2013新兴技术最佳演示奖，2019年和2020年连续两年计算机视觉顶级会议CVPR最佳论文列表（入选率分别为0.8%和0.4%），入选2021腾讯AI Lab犀牛鸟专项研究计划，他的团队于2018年11月获得IEEE ICDM 全球气象挑战赛冠军（参赛队伍1700多）。

研究方向：

计算机视觉、计算机图形学、人机交互、医学图像处理、机器学习

概要：

基于深度学习的方法在可视媒体，如图像、视频和三维数据的理解任务中已经被广泛应用。但目前大部分的方法均利用隐式表达来对内容进行建模，使得算法缺乏可解释性以及难以控制和训练。主讲人的研究团队近期一直在尝试如何将显式和隐式表达进行有效结合从而更好进行建模。该报告将通过介绍他们近期的三个工作尝试（分别是在图像、视频理解以及三维建模方面）给出一些发现和观点。

Swipe up to view

Dr. Xiaoguang Hanis now an Assistant Professor of the School of Science and Engineering. He received his Ph.D. degree in computer science from The University of Hong Kong (2013-2017), his M.S. degree in applied mathematics from Zhejiang University (2009-2011) and his B.S. degree in math from Nanjing University of Aeronautics and Astronautics. He also spent 2 years (2011-2013) in City University of Hong Kong as a research associate. In Sep 2017, he joinedCUHK-Shenzhen and Shenzhen Research Institute of Big Data (SRIBD).

He already published more than 40 papers at the leading conferences and journals, including SIGGRAPH, CVPR, ICCV,ECCV,IEEE TPAMI, ACM TOG etc. Histwoworkshavebeen respectively selected as the CVPR 2019 and CVPR 2020 best paper finalist (acceptance rate is around 0.8%and 0.4%), his team received the ChinaGraph Best Dataset Award, his team also got the 1st place of IEEE ICDM Global A.I. Challenge on Meteorology (out of 1700 teams).

ResearchAreas:

computer vision, computer graphics, human-computer interaction, medical image analysis and machine learning

Abstract:

Recently, deep learning-based methods dominate all of the works on image, video and 3D understanding. They found that all existing methods utilized implicit representation for visual modeling. In this talk, the speaker will introduce recent three works to show their findings that combing explicit and implicit representation can benefit existing methods. The works include image instance segmentation, tumor segmentation from DSA video input and 3D modeling from 2D sketching.

吴保元

数据科学学院副教授

深圳市大数据研究院大数据安全计算实验室主任

Neurocomputing期刊编委

吴保元博士现任香港中文大学（深圳）数据科学学院副教授，深圳市大数据研究院大数据安全计算实验室主任。2014年获得中国科学院自动化研究所模式识别国家重点实验室模式识别与智能系统博士学位。2016年11月至2018年12月担任腾讯AI Lab高级研究员，2019年1月至2020年8月担任专家研究员。他在人工智能安全与隐私、机器学习、计算机视觉、优化等方向上做出了多项出色工作，在人工智能的顶级期刊和会议上发表论文40多篇，包括TPAMI、IJCV、CVPR、ICCV、ECCV、ICLR、AAAI等，并曾入选人工智能顶级会议CVPR 2019最佳论文候选名单。其担任人工智能领域国际期刊Neurocomputing编委、国际会议ICLR 2022、ICIG 2021 领域主席、国际会议AAAI 2021、IJCAI 2020/2021高级程序委员、中国计算机学会、中国自动化学会多个专业委员会委员。作为项目负责人承担国家自然科学基金面上项目1项，腾讯研究专项基金2项。

研究领域：

机器学习、计算机视觉和优化

概要：

本次演讲中，讲者将首先简单介绍大数据安全计算实验室，以及人工智能安全的定义和研究范围，随后将重点介绍现实场景下的人工智能安全问题的最新研究进展。

在现实场景下，很多人工智能系统以API接口的形式对外提供服务，比如云提供商提供的人脸识别、图像识别等API服务。对于攻击者来说，只能通过查询API获取预测结果，而无法获知API接口背后的模型参数、结构以及训练数据等信息。对于服务提供者来说，既需要为每一次查询提供稳定而准确的预测，又需要防范未知恶意攻击所带来的安全风险，这种情形称为“现实场景下的黑盒对抗攻击与防御”。主讲人将首先介绍两个黑盒对抗攻击方法，包括：业内第一个以黑盒形式成功欺骗人脸识别API的决策式攻击算法，其基本思想利用历史查询来加速搜索过程；另一个是基于返回分数的黑盒攻击方法，通过条件流模型对对抗扰动的概率分布进行建模，该概率分布被用作进化算法中的采样分布，大幅度提高了黑盒攻击效率和成功率。他还将介绍一种轻量级的黑盒防御方法，被称为随机噪声防御(RND)，其基本思想是为每次查询添加适当的高斯噪声来干扰攻击过程。他们为RND对查询式攻击的防御有效性进行了理论分析和实验验证。RND 还可以和现有的防御方法结合，进一步增强模型的安全性。

Swipe up to view

Dr. Baoyuan Wu is an Associate Professor of the School of Data Science. He is also the director of the Secure Computing Lab of Big Data, Shenzhen Research Institute of Big Data. He received the PhD degree from the National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences in 2014. From November 2016 to August 2020, he was a Senior and Principal Researcher at Tencent AI lab. His research interests are AI security and privacy, machine learning, computer vision and optimization. He has published 40+ top-tier conference and journal papers, including TPAMI, IJCV, CVPR, ICCV, ECCV, AAAI, and one paper was selected as the Best Paper Finalist of CVPR 2019. He is currently serving as an Associate Editor of Neurocomputing, Area Chair of ICLR 2022 and ICIG 2021, Senior Program Committee Member of AAAI 2021 and IJCAI 2020/2021, Task Force Member of CCF and CAA. He is the principal investigator of General Program of National Natural Science Foundation of China and Tencent Rhino Bird Special Research Fund.

ResearchAreas:

machinelearning,computervision andoptimization

Abstract:

In this talk, the speaker will first briefly introduce the Security Computing Lab of Big Data (SCLBD), and the definition and research scope of AI security, and then will focus on the latest research progress of AI security issues in real scenarios.

Many AI systems provide services to the public in the form of API interfaces, such as face recognition, image recognition and other API services provided by cloud providers. For attackers, they can only obtain the prediction results by querying the API, but they cannot get information about the model parameters, structure and training data behind the API interface. For the service provider, it needs to provide stable and accurate predictions for each query, but also needs to prevent the security risk caused by unknown malicious attacks. They refer to this scenario as "black-box adversarial attacks and defenses in real-world scenarios". The speaker will first introduce two black-box attack methods, including: the first decision-based attack algorithm that successfully fools face recognition APIs in the black box setting. The basic idea is to use historical queries to speed up the search process; and another black-box attack method based on return scores, which models the probability distribution of attack perturbations through a conditional flow model, which is used as the sampling distribution in an evolutionary algorithm, and substantially improves the efficiency and success rate of black-box attacks. He will also introduce a lightweight black-box defense method, called Random Noise Defense (RND). The basic idea is to add an appropriate amount of Gaussian noise to each query to disrupt the attack process. They perform a theoretical analysis and experimental validation of the effectiveness of the defense against query-based attacks for RND, which can also be combined with existing defense methods to further enhance the security of the model.

End

传讯及公共关系处出品

排版：房奕珺（2018级经管学院、思廷书院）

CUHK-Shenzhen

香港中文大学（深圳）

结合传统与现代 融会中国与西方