[{"data":1,"prerenderedAt":385},["ShallowReactive",2],{"content-query-sz9MTs91iL":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":379,"_id":380,"_source":381,"_file":382,"_stem":383,"_extension":384},"/technology-blogs/zh/3112","zh",false,"","论文精讲 | 基于昇思MindSpore的多维度公平性联邦学习","作者：李锐锋 ｜来源：知乎","2024-05-13","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/05/17/e7eab09ab5eb45029db4937f0e82b5d5.png","technology-blogs",{"type":14,"children":15,"toc":376},"root",[16,24,43,48,53,58,63,68,79,84,93,98,117,125,133,141,146,151,159,167,179,190,198,206,211,218,223,228,233,241,249,254,261,266,271,278,283,290,295,300,308,316,321,326,331,339,366],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"论文精讲-基于昇思mindspore的多维度公平性联邦学习",[22],{"type":23,"value":8},"text",{"type":17,"tag":25,"props":26,"children":27},"p",{},[28,30,36,38],{"type":23,"value":29},"**作者：**",{"type":17,"tag":31,"props":32,"children":33},"strong",{},[34],{"type":23,"value":35},"李锐锋",{"type":23,"value":37}," ｜",{"type":17,"tag":31,"props":39,"children":40},{},[41],{"type":23,"value":42},"来源：知乎",{"type":17,"tag":25,"props":44,"children":45},{},[46],{"type":23,"value":47},"论文标题",{"type":17,"tag":25,"props":49,"children":50},{},[51],{"type":23,"value":52},"Multi-Dimensional Fair Federated Learning",{"type":17,"tag":25,"props":54,"children":55},{},[56],{"type":23,"value":57},"论文来源",{"type":17,"tag":25,"props":59,"children":60},{},[61],{"type":23,"value":62},"AAAI 2024",{"type":17,"tag":25,"props":64,"children":65},{},[66],{"type":23,"value":67},"论文链接",{"type":17,"tag":25,"props":69,"children":70},{},[71],{"type":17,"tag":72,"props":73,"children":77},"a",{"href":74,"rel":75},"https://ojs.aaai.org/index.php/AAAI/article/view/29430",[76],"nofollow",[78],{"type":23,"value":74},{"type":17,"tag":25,"props":80,"children":81},{},[82],{"type":23,"value":83},"代码链接",{"type":17,"tag":25,"props":85,"children":86},{},[87],{"type":17,"tag":72,"props":88,"children":91},{"href":89,"rel":90},"https://www.sdu-idea.cn/codes.php?name=mFairFL",[76],[92],{"type":23,"value":89},{"type":17,"tag":25,"props":94,"children":95},{},[96],{"type":23,"value":97},"昇思MindSpore作为开源的AI框架，为产学研和开发人员带来端边云全场景协同、极简开发、极致性能、安全可信的体验，支持超大规模AI预训练，自2020年3月28日开源来下载量已超过7百万。昇思MindSpore已支持1000+篇AI顶会论文，走入Top100+高校教学，通过HMS在5000+App上商用，拥有数量众多的开发者，在AI计算中心、智能制造、金融、云、无线、数通、能源、消费者1+8+N、智能汽车等端边云车全场景广泛应用，是Gitee指数最高的开源软件。欢迎大家参与开源贡献、套件、模型众智、行业创新与应用、算法创新、学术合作、AI书籍合作等，贡献您在云侧、端侧、边侧以及安全领域的应用案例。",{"type":17,"tag":25,"props":99,"children":100},{},[101,103,108,110,115],{"type":23,"value":102},"在科技界、学术界和工业界对昇思MindSpore的广泛支持下，基于昇思MindSpore的AI论文2023年在所有AI框架中占比7%，连续两年进入全球第二，感谢CAAI和各位高校老师支持，我们一起继续努力做好AI科研创新。昇思MindSpore社区支持顶级会议论文研究，持续构建原创AI成果。我会不定期挑选一些优秀的论文来推送和解读，希望更多的产学研专家跟昇思MindSpore合作，一起推动原创AI研究，昇思MindSpore社区会持续支撑好AI创新和AI应用，本文是昇思MindSpore AI顶会论文系列第34篇，我选择了来自山东大学软件学院的",{"type":17,"tag":31,"props":104,"children":105},{},[106],{"type":23,"value":107},"余国先",{"type":23,"value":109},"和",{"type":17,"tag":31,"props":111,"children":112},{},[113],{"type":23,"value":114},"王峻",{"type":23,"value":116},"老师团队发表于AAAI2024的一篇论文解读，感谢各位专家教授同学的投稿，更多精彩的论文精读文章和开源代码实现请扫下方二维码访问Models。",{"type":17,"tag":25,"props":118,"children":119},{},[120],{"type":17,"tag":121,"props":122,"children":124},"img",{"alt":7,"src":123},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/05/17/be066a4a74d1412d82aee3d0132f2fd5.png",[],{"type":17,"tag":25,"props":126,"children":127},{},[128],{"type":17,"tag":31,"props":129,"children":130},{},[131],{"type":23,"value":132},"01",{"type":17,"tag":25,"props":134,"children":135},{},[136],{"type":17,"tag":31,"props":137,"children":138},{},[139],{"type":23,"value":140},"研究背景",{"type":17,"tag":25,"props":142,"children":143},{},[144],{"type":23,"value":145},"算法的公平性问题越来越受到人们的关注，其要求算法决策公平地对待不同敏感群体（如不同性别的群体）。现有公平性算法要求能够直接访问所有的训练数据，然而数据集的所有权往往属于不同的机构，出于隐私保护的考虑，使得它们无法共享。联邦学习是一种分布式机器学习解决方案，允许在不共享数据的情况下，基于所有参与者的数据训练一个全局模型。",{"type":17,"tag":25,"props":147,"children":148},{},[149],{"type":23,"value":150},"然而，目前在数据离散场景下的方法缺乏协作公平性。一方面，全局模型对于各个客户端在预测性能上存在较大差异（客户端公平）；另一方面，全局模型对受保护群体的预测存在歧视（群体公平）。为了解决这些问题，作者提出了一种多维度公平性联邦学习方法(mFairFL)，其旨在确保学习的全局模型同时实现群体公平和客户端公平。mFairFL利用微分乘子构建具有公平性约束的经验风险最小化的优化目标，并在中央服务器聚合局部模型之前，其首先检测这些客户端梯度是否有冲突，然后迭代的调整这些有冲突的梯度的大小和方向，以此缓解客户端梯度冲突。本文算法可以按照昇思MindSpore官方文档案例以及提供的代码进行实现，在不同程度的客户端数据异质性设置下进行的实验及分析验证了所提出方法的有效性。",{"type":17,"tag":25,"props":152,"children":153},{},[154],{"type":17,"tag":31,"props":155,"children":156},{},[157],{"type":23,"value":158},"02",{"type":17,"tag":25,"props":160,"children":161},{},[162],{"type":17,"tag":31,"props":163,"children":164},{},[165],{"type":23,"value":166},"团队介绍",{"type":17,"tag":25,"props":168,"children":169},{},[170,172,177],{"type":23,"value":171},"论文第一作者",{"type":17,"tag":31,"props":173,"children":174},{},[175],{"type":23,"value":176},"苏聪",{"type":23,"value":178},"目前就读于山东大学软件学院（2022-至今），硕士研究生，主要研究方向包括公平性机器学习和因果学习，导师为余国先教授。",{"type":17,"tag":25,"props":180,"children":181},{},[182,184,188],{"type":23,"value":183},"第一作者导师",{"type":17,"tag":31,"props":185,"children":186},{},[187],{"type":23,"value":107},{"type":23,"value":189},"，山东大学软件学院教授，博导，国家重点研发青年科学家项目负责人，泰山学者青年专家，齐鲁青年学者，小米青年学者。主要从事人工智能，机器学习，数据挖掘及其在生物医学数据分析中的应用研究。在国内外主流期刊和会议(TKDE、TNNLS、TDSC、NAR、Bioinformatics、KDD、ICDE、AAAI、IJCAI、中国科学-信息科学、计算机学报等)发表论文100余篇,获得重庆市自然科学奖1项，山东计算机学会自然科学一等奖1项，全球前2%顶尖科学家（2020-2023），担任TNNLS、Neurocomputing、Interdisciplinary Sciences: Computational Life Sciences、Frontiers in Genetics等人工智能和生物信息领域SCI期刊副主编和AAAI/IJCAI/KDD 等主流人工智能会议领域主席、高级程序委员等。",{"type":17,"tag":25,"props":191,"children":192},{},[193],{"type":17,"tag":31,"props":194,"children":195},{},[196],{"type":23,"value":197},"03",{"type":17,"tag":25,"props":199,"children":200},{},[201],{"type":17,"tag":31,"props":202,"children":203},{},[204],{"type":23,"value":205},"论文简介",{"type":17,"tag":25,"props":207,"children":208},{},[209],{"type":23,"value":210},"确保模型决策公平的问题直接影响到社会和公众对人工智能技术的信任。所谓的模型公平性是指模型决策不存在基于个人或群体的内在或后天特征（例如性别、种族和年龄等）的任何歧视。绝大部分的公平性算法要求可以直接访问完整的数据，然而，这些数据由不同机构所持有，出于对数据隐私保护的要求，使得它们无法共享。本文主要针对数据分散存储场景下，跨机构间难以有效地协同训练公平性模型的问题，着眼于如何在不泄露数据隐私的前提下，训练一个同时满足群体公平（Group Fairness）和客户端公平（Client Fairness）的模型。由于各客户端之间数据分布异质性的问题，以往的联邦式公平性算法以直接加权聚合本地局部模型的方式不仅会与数据集中式训练处的模型表现出显著差异，而且还会导致全局模型在更新过程中过度偏爱某些客户端，进而对其他客户端产生不利影响，从而产生所谓的客户端公平性问题。",{"type":17,"tag":25,"props":212,"children":213},{},[214],{"type":17,"tag":121,"props":215,"children":217},{"alt":7,"src":216},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/05/17/3fa13e21ab1b4db99554f7b4ef80cbd5.png",[],{"type":17,"tag":25,"props":219,"children":220},{},[221],{"type":23,"value":222},"图1 本文方法的模型框架图",{"type":17,"tag":25,"props":224,"children":225},{},[226],{"type":23,"value":227},"为此，本文提出了一种新的多维度公平性联邦学习方法mFairFL，其框架结构如图1所示。具体来说，为了从分散的数据中训练一个满足群体公平性约束的全局模型，mFairFL将群体公平性约束纳入到模型训练过程中，并且通过引入拉格朗日乘子将其转换成无约束优化问题。该无约束目标可以通过使用梯度下降或上升的方式进行优化。通过这种方式，各客户端根据本地数据计算本地局部模型统计信息，并将其上传给服务器，包括局部模型损失，模型梯度。",{"type":17,"tag":25,"props":229,"children":230},{},[231],{"type":23,"value":232},"由于客户端数据的异质性，这些客户端局部模型的梯度可能存在冲突，特别是当冲突梯度的大小存在较大差异时，全局模型的更新会以损害其在某些客户端上的性能表现为代价，偏向另一些客户端。因此，在服务器聚合阶段，为了缓解客户端梯度冲突产生的不利影响，实现客户端公平和群体公平，mFairFL首先根据客户端损失的大小按升序排列客户端梯度，其表示了将每个客户端梯度作为参考梯度投影目标的序列。然后按照该序列规定的投影顺序，mFairFL迭代地调整客户端的梯度，使其符合期望的梯度相似性标准。最后，服务器对调整后的客户端梯度进行聚合以更新全局模型，并将更新后的全局模型返回给各客户端。通过局部与全局的迭代循环优化，最终可以得到一个性能优异且满足公平性要求的全局模型。",{"type":17,"tag":25,"props":234,"children":235},{},[236],{"type":17,"tag":31,"props":237,"children":238},{},[239],{"type":23,"value":240},"04",{"type":17,"tag":25,"props":242,"children":243},{},[244],{"type":17,"tag":31,"props":245,"children":246},{},[247],{"type":23,"value":248},"实验结果",{"type":17,"tag":25,"props":250,"children":251},{},[252],{"type":23,"value":253},"在客户端数据异构设置下，作者比较了mFairFL与目前最先进的考虑群体公平的联邦学习方法在三个真实数据上的性能，并进一步对实验结果进行了分析研究。如表1所示，mFairFL能够有效的提高模型公平性，实现类似在数据集中式设置下的公平性性能，且最先进的方法（SOTA）相比，mFairFL能够更加有效地权衡模型公平性和精确度。这证实了mFairFL在数据分散设置下巧妙地为敏感群体训练公平性模型方面的有效性。",{"type":17,"tag":25,"props":255,"children":256},{},[257],{"type":17,"tag":121,"props":258,"children":260},{"alt":7,"src":259},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/05/17/b3f8c7cd027a48d58173dd0668ac7507.png",[],{"type":17,"tag":25,"props":262,"children":263},{},[264],{"type":23,"value":265},"表1 数据低异质下mFairFL与考虑群体公平的联邦学习方法比较的实验结果",{"type":17,"tag":25,"props":267,"children":268},{},[269],{"type":23,"value":270},"作者还在相同的实验设置下比较了mFairFL与目前最先进的考虑客户端公平的联邦学习方法的性能。表2中的结果显示，与SOTA方法相比，mFairFL实现了最优的客户端公平表现，同时实现了与他们相当的预测准确性。该实验反应了mFairFL在缓解客户端梯度冲突的有效性，确保全局模型的更新对各客户端有利。",{"type":17,"tag":25,"props":272,"children":273},{},[274],{"type":17,"tag":121,"props":275,"children":277},{"alt":7,"src":276},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/05/17/69a70c3c567c4379978c8082a327f8d3.png",[],{"type":17,"tag":25,"props":279,"children":280},{},[281],{"type":23,"value":282},"表2 数据高异质下mFairFL与考虑群体公平的联邦学习方法比较的实验结果",{"type":17,"tag":25,"props":284,"children":285},{},[286],{"type":17,"tag":121,"props":287,"children":289},{"alt":7,"src":288},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/05/17/3627bb144e2748d0b738c705c4ebc30a.png",[],{"type":17,"tag":25,"props":291,"children":292},{},[293],{"type":23,"value":294},"表3 数据高异质下mFairFL与考虑群体公平的联邦学习方法比较的实验结果",{"type":17,"tag":25,"props":296,"children":297},{},[298],{"type":23,"value":299},"这些实验结果和理论分析标志着mFairFL在数据分散场景下部署应用具有明显的潜力。",{"type":17,"tag":25,"props":301,"children":302},{},[303],{"type":17,"tag":31,"props":304,"children":305},{},[306],{"type":23,"value":307},"05",{"type":17,"tag":25,"props":309,"children":310},{},[311],{"type":17,"tag":31,"props":312,"children":313},{},[314],{"type":23,"value":315},"总结与展望",{"type":17,"tag":25,"props":317,"children":318},{},[319],{"type":23,"value":320},"在联邦学习领域中，群体公平和客户端公平是联邦学习领域中两个重要的亟需解决的公平性问题。本文提出了一种有效的方法mFairFL。mFairFL将优化问题定义为一个具有群体公平约束的极大极小问题，并在全局模型训练时调整有冲突的客户端梯度的大小和方向，以同时实现群体公平和客户端公平。理论研究和实证结果都验证了mFairFL的有效性。",{"type":17,"tag":25,"props":322,"children":323},{},[324],{"type":23,"value":325},"在使用昇思MindSpore复现本文的算法时，我们发现昇思MindSpore框架将模型参数和梯度相分离，且昇思MindSpore框架的函数自动微分暴露的接口更底层，对于梯度的处理更加灵活。此外，官方的昇思MindSpore教程文档与样例可帮助入门，完成相关算法的设计与实现，同时还包含了与其他深度学习框架的API对照文档，能够帮助实现模型的迁移。尽管如此，仍需要更多的开发者和研究者参与到昇思MindSpore社区建设中。我们期望更多开发者和研究者加入进来，通过分享个人的经验和技巧，共同促进知识交流，助力解决彼此在开发过程中遇到的难题。",{"type":17,"tag":25,"props":327,"children":328},{},[329],{"type":23,"value":330},"往期回顾",{"type":17,"tag":18,"props":332,"children":333},{"id":7},[334],{"type":17,"tag":72,"props":335,"children":338},{"href":336,"rel":337},"http://mp.weixin.qq.com/s?__biz=MzkxMTM2MjMzNg==&mid=2247614420&idx=1&sn=8d9c49e82b9b330a1b7ff0fa3a5a24ac&chksm=c11e2f9bf669a68d9f500863e9c15fc324aa8d37ddff506d6c8b71ff4bc4ac391aacfc55f076&scene=21#wechat_redirect",[76],[],{"type":17,"tag":25,"props":340,"children":341},{},[342,349,351,358,359],{"type":17,"tag":72,"props":343,"children":346},{"href":344,"rel":345},"http://mp.weixin.qq.com/s?__biz=MzkxMTM2MjMzNg==&mid=2247619744&idx=1&sn=753cb514c696f09927f6bc47e01bf929&chksm=c11e04eff6698df903e5eaafdf3ae26a18210fe4f8a3f0e56c2253ca31bacb54d25c298e1b79&scene=21#wechat_redirect",[76],[347],{"type":23,"value":348},"论文精讲 | 基于昇思MindSpore具有可解释自适应优化的联邦因果发现",{"type":23,"value":350}," ",{"type":17,"tag":72,"props":352,"children":355},{"href":353,"rel":354},"http://mp.weixin.qq.com/s?__biz=MzkxMTM2MjMzNg==&mid=2247619475&idx=1&sn=58bf1fd0e26efd38da3ddf7630c5b8e3&chksm=c11e03dcf6698aca03418f905d2052736b9d371e82bfd77ba7834938255ecdc84b4ff3042676&scene=21#wechat_redirect",[76],[356],{"type":23,"value":357},"论文精讲 | 基于昇思MindSpore的多粒度因果结构学习有效提升因果关系发现的准确率",{"type":23,"value":350},{"type":17,"tag":72,"props":360,"children":363},{"href":361,"rel":362},"http://mp.weixin.qq.com/s?__biz=MzkxMTM2MjMzNg==&mid=2247616069&idx=1&sn=40a89de0df0ee584097992d15f804809&chksm=c11e160af6699f1cc5530e86f404880fcbc5e306ea6ba5ea61a800de8e788cc58e4cd5e2803c&scene=21#wechat_redirect",[76],[364],{"type":23,"value":365},"论文精讲 | 基于昇思MindSpore评测的FLAG3D——自然语言引导的三维健身动作数据集",{"type":17,"tag":25,"props":367,"children":368},{},[369],{"type":17,"tag":72,"props":370,"children":373},{"href":371,"rel":372},"http://mp.weixin.qq.com/s?__biz=MzkxMTM2MjMzNg==&mid=2247615411&idx=1&sn=3473cb214717cdbb546a91d3bc905756&chksm=c11e13fcf6699aeabe0978a6ff8273dc41e50adf9c34c3597a494c51b71d2eabf9aa1cd53465&scene=21#wechat_redirect",[76],[374],{"type":23,"value":375},"论文精讲 | 基于昇思MindSpore的加速对抗训练算法，可显著减少训练时间",{"title":7,"searchDepth":377,"depth":377,"links":378},4,[],"markdown","content:technology-blogs:zh:3112.md","content","technology-blogs/zh/3112.md","technology-blogs/zh/3112","md",1776506126078]