20210630-17 总第243期 面向伦理与隐私保护的智能技术

报告嘉宾：王帝 (King Abdullah University of Science and Technology)

报告题目：Recent Developments of Differential Privacy in Classical and Modern Machine Learning

报告嘉宾：许正 (Google Research)

报告题目：Practical and Private Federated Learning

Panel嘉宾：

王帝 (King Abdullah University of Science and Technology)、许正 (Google Research)、于涵 (Nanyang Technological University)、崔鹏 (清华大学)、张弘扬 (University of Waterloo & Vector Institute)

Panel议题：

1. 如何兼顾人工智能应用与隐私保护？

2. 对于人工智能伦理的研究，会如何帮助Artificial General Intelligence的实现？

3. 人工智能算法独立做出判断及决策。那么在研发的过程中，如何确保人工智能的这种判断及决策能够符合人类的伦理准则？

4. 联邦学习方法中，使用梯度信息是否足够安全？梯度信息的泄露，可能进而导致部分数据的泄露？

5. 联邦学习有很多应用场景，例如医疗、金融、保险等。在这些场景的应用过程中，有哪些困难或挑战？以及可能的应对方法？

6. 公平 (fairness), 隐私 (privacy), 可解释性 (explanability), 鲁棒性 (robustness)等都属于可信机器学习 (trustworthy machine learning)的范畴， 如何将他们联系起来？或者说他们之间是如何互相促进的？

*欢迎大家在下方留言提出主题相关问题，主持人和panel嘉宾会从中选择若干热度高的问题加入panel议题！

报告嘉宾：王帝 (King Abdullah University of Science and Technology)

报告时间：2021年06月30日 (星期三)晚上21:30 (北京时间)

报告题目：Recent Developments of Differential Privacy in Classical and Modern Machine Learning

报告人简介：

Di Wang is currently an Assistant Professor at the King Abdullah University of Science and Technology (KAUST). Before that, he got his PhD degree in the Computer Science and Engineering at the State University of New York (SUNY) at Buffalo. And he obtained his BS and MS degrees in mathematics from Shandong University and the University of Western Ontario, respectively. During his PhD studies, he has been invited as a visiting student to the University of California, Berkeley, Harvard University, and Boston University. His research areas include differentially private machine learning, interpretable machine learning, robust estimation and optimization. He has received the SEAS Dean’s Graduate Achievement Award and the Best CSE Graduate Research Award from SUNY Buffalo.

个人主页：

https://shao3wangdi.github.io

报告摘要：

Recent research showed that most of the existing learning models are vulnerable to various privacy attacks. Thus, a major challenge facing the machine learning community is how to learn effectively from sensitive data. An effective way for this problem is to enforce differential privacy during the learning process. As a rigorous scheme for privacy preserving, Differential Privacy (DP) has now become a standard for private data analysis.  Despite its rapid development in theory, DP's adoption to the machine learning community remains slow due to various challenges from the data, the privacy models and the learning tasks. In this talk, I will use the Empirical Risk Minimization (ERM) problem as an example and show its recent developments and challenges in classical and modern machine learning.

参考文献：

[1] Di Wang, Marco Gaboardi, Adam Smith, and Jinhui Xu. "Empirical Risk Minimization in the Non-interactive Local Model of Differential Privacy." Journal of machine learning research 21, no. 200 (2020).

[2] Di Wang, and Jinhui Xu. "On Sparse Linear Regression in the Local Differential Privacy Model." IEEE Transactions on Information Theory (2020).

[3] Di Wang, Hanshen Xiao, Srini Devadas, and Jinhui Xu. "On Differentially Private Stochastic Optimization with Heavy-tailed Data" ICML 2020.

[4] Di Wang, Changyou Chen, and Jinhui Xu. "Differentially private empirical risk minimization with non-convex loss functions." ICML 2019.

[5] Di Wang, Marco Gaboardi and Jinhui Xu. Empirical Risk Minimization in Non-interactive Local Differential Privacy Revisited. NIPS 2018.

报告嘉宾：许正 (Google Research)

报告时间：2021年06月30日 (星期三)晚上22:00 (北京时间)

报告题目：Practical and Private Federated Learning

报告人简介：

Zheng Xu is a research scientist working on federated learning at Google. He got his Ph.D. in optimization and machine learning from University of Maryland, College Park. Before that, he got his master's and bachelor's degree from University of Science and Technology of China. During the studies, Zheng has interned and collaborated with researchers from Apple, Adobe, Honda, Amazon, IBM, MSRA and NTU. His papers have received best student paper awards at workshops and conferences like ICML PADL Workshop.

个人主页：

https://sites.google.com/site/xuzhustc/

报告摘要：

Federated learning (FL) is an emerging paradigm to perform distributed model training on decentralized (edge) clients under the orchestration of a central server, while keeping private data on the client devices. Practical FL problems often emphasize communication efficiency, data heterogeneity, compatibility with privacy and system requirements, and other constraints that are not primary considerations in other (distributed) problem settings. In this talk, we first consider training FL models with provable differential privacy guarantees through a method called DP-FTRL. DP-FTRL avoids sampling and achieves similar and sometimes better privacy-utilty trade-offs as DP-SGD. We then briefly discuss heterogeneity and the resulting consistency problem and possible optimization solvers.

参考文献：

[1] Peter Kairouz, Brendan McMahan, Shuang Song, Om Thakkar, Abhradeep Thakurta, and Zheng Xu. Practical and Private (Deep) Learning without Sampling or Shuffling. ICML 2021.

[2] Jianyu Wang, Zheng Xu, Zachary Garrett, Zachary Charles, Luyang Liu, and Gauri Joshi. Exploiting Local Adaptivity in Federated Learning. arxiv 2021.

[3] Peter Kairouz, Brendan McMahan, and more than 50 authors including Zheng Xu. Advances and Open Problems in Federated Learning. arxiv 2019.

Panel嘉宾：于涵 (Nanyang Technological University)

嘉宾简介：

于涵博士现任新加坡南洋理工大学计算机科学与工程系南洋助理教授。2015至2018年间，他在南洋理工大学担任李光耀研究员。他于2007取得南洋理工大学计算机工程系一等荣誉学士学位，并于2014年取得南洋理工大学计算机工程系博士学位。2007至2008年间，他在惠普新加坡担任嵌入式软件工程师职务。他的研究主要致力于联邦学习及其应用。他的科研成果在AI国际学术会议及期刊上发表多篇论文，并多次在AAAI和IJCAI等前沿国际会议及期刊中获奖。

个人主页：

http://hanyu.sg/

Panel嘉宾：崔鹏 (清华大学)

嘉宾简介：

崔鹏，清华大学计算机系长聘副教授，2010年于清华大学获得博士学位。研究兴趣包括大数据环境下的因果推理与稳定预测、网络表征学习，及其在金融科技、智慧医疗及社交网络等场景中的应用。已在数据挖掘及多媒体领域顶级国际期刊和会议上发表论文百余篇，并先后获得7项国际会议及期刊最佳论文奖。目前担任IEEE TKDE、IEEE TBD、ACM TOMM、ACM TIST等国际顶级期刊的编委。获得中国计算机学会青年科学家奖，国际计算机协会 (ACM)中国新星奖，并入选中国科协首届青年人才托举计划。获得国家自然科学二等奖、教育部自然科学一等奖、北京市科技进步一等奖、中国电子学会自然科学一等奖。入选中组部万人计划青年拔尖人才，并当选为中国科协全国委员会委员。

个人主页：

http://pengcui.thumedialab.com

Panel嘉宾：张弘扬 (University of Waterloo & Vector Institute)

嘉宾简介：

Dr. Zhang is an assistant professor at University of Waterloo, David R. Cheriton School of Computer Science, and affiliated with Vector Institute. He completed his Ph.D. from 2015-2019 in Carnegie Mellon University, Machine Learning Department. After that, he was a Postdoctoral Research Associate at Toyota Technological Institute at Chicago. Dr. Zhang’s research interests include trustworthy ML and self-supervised learning. He also won the championship in multiple competitions, including CVPR’21 Security AI Challenger: Unrestricted Adversarial Attacks on ImageNet (1st out of 1,559 teams), NeurIPS’18 Adversarial Vision Challenge (1st out of 400 teams), and Unrestricted Adversarial Example Challenge. His work has been widely recognized by DeepMind, Google, Microsoft, etc. Besides practical contributions, Dr. Zhang is also interested in theory and has published 1 book and more than 30 papers in SODA, COLT, ITCS, ICALP, JMLR, Proceedings of IEEE, etc.

个人主页：

https://hongyanz.github.io

主持人：徐畅 (The University of Sydney)

主持人简介：

徐畅，悉尼大学高级讲师，澳大利亚研究委员会优秀青年基金获得者 (ARC DECRA)。主要关注机器学习算法及其在计算机视觉中的应用，如多视角学习、多标签学习、对抗机器学习、神经网络结构设计与搜索等。在人工智能领域重要国际会议和期刊发表论文100余篇，包括ICML、NIPS、CVPR、IJCAI、AAAI、IEEE T-PAMI和IEEE T-IP等。

个人主页：

http://changxu.xyz

21-17期VALSE在线学术报告参与方式：

长按或扫描下方二维码，关注“VALSE”微信公众号 (valse_wechat)，后台回复“17期”，获取直播地址。

特别鸣谢本次Webinar主要组织者：

主办AC：徐畅 (The University of Sydney)

协办AC：张弘扬 (University of Waterloo & Vector Institute)