联邦学习框架下的数据安全与利用合规路径
网络安全与数据治理 6期
孙绮雯
(清华大学法学院，北京100084)
摘要： 日趋严格的个人信息保护相关法律法规，在保护个人隐私的同时，增加了企业数据流通合规的难度和成本。在联邦学习框架中，数据不动模型动的隐私保护设计以技术促进法律的遵守，是打破数据孤岛壁垒、促进隐私保护前提下数据融合协作创新的可能解。将合法原则、数据最小化原则与目的限制原则嵌入到系统开发的技术中，联邦学习分布式协作框架以局部模型更新参数代替本地原始个人数据上传，实现数据本地训练存储，达到可用不可见的个人信息保护效果。由于潜在的网络安全攻击以及机器学习算法黑箱的固有缺陷，联邦学习仍然面临着质量原则、公正原则与透明原则的挑战。联邦学习不是规避合规义务的手段，而是减少个人信息合规风险的可行技术措施，使用时仍然存在需要履行的个人信息保护义务，数据权属与责任分配的确定需要综合考量各参与方角色和个人信息处理者类型。
中圖分類號：D922.174
文獻標識碼：A
DOI:10.19358/j.issn.2097-1788.2023.06.004
引用格式：孫綺雯.聯(lián)邦學(xué)習(xí)框架下的數(shù)據(jù)安全與利用合規(guī)路徑［J］.網(wǎng)絡(luò)安全與數(shù)據(jù)治理，2023，42（6）：21-29.
文獻標識碼：A
DOI:10.19358/j.issn.2097-1788.2023.06.004
引用格式：孫綺雯.聯(lián)邦學(xué)習(xí)框架下的數(shù)據(jù)安全與利用合規(guī)路徑［J］.網(wǎng)絡(luò)安全與數(shù)據(jù)治理，2023，42（6）：21-29.
Data security and utilization compliance path under the federated learning framework
Sun Qiwen
(School of Law, Tsinghua University, Beijing 100084, China）
Abstract： The increasingly stringent laws and regulations related to personal information protection have increased the difficulty and cost of compliance in data circulation of enterprises while protecting personal privacy. Under the framework of federated learning, the privacy protection design that does not transmit the original data but only transmits the model uses technology to promote legal compliance, which can be a possible solution for data fusion and collaborative innovation under the premise of breaking the barriers of data isolation and promoting privacy protection. The legal principles, data minimization principle and purpose limitation principle, are embedded into the technical process of the system development. The distributed collaborative framework of federated learning uploads the updated parameters of the local model instead of original personal data, realizing local training and storage of data, and achieving such a great personal information protection effect that data can be utilizable while at the same time invisible. Due to potential network security attacks and inherent defects of machine learning algorithms black box, federated learning still faces the challenges of the principles of quality, fairness, and transparency. Federated learning is not a way to evade compliance obligations, but a feasible technical measure to reduce compliance risks of personal information. There still exist personal information protection obligations to be fulfilled when using federated learning framework. The determination of data ownership and responsibility allocation requires comprehensively consideration of the roles of each participant and the types of personal information processors.
Key words : federated learning; personal information protection; isolated data island; network security attack; collaborate and share
0 引言
當前人工智能發(fā)展面臨數(shù)據(jù)孤島現(xiàn)象與數(shù)據(jù)融合需求的矛盾，聯(lián)邦學(xué)習(xí)有助于破解數(shù)據(jù)協(xié)作創(chuàng)新與數(shù)據(jù)隱私保護的困境。作為基于設(shè)計隱私的分布式協(xié)作模型，聯(lián)邦學(xué)習(xí)可以在保護個人信息的前提下，使得跨組織、跨設(shè)備、跨區(qū)域的不同特征維度數(shù)據(jù)合規(guī)共享、流通、融合。在聯(lián)邦學(xué)習(xí)框架中還可以結(jié)合使用多種隱私計算技術(shù)，如多方安全計算、同態(tài)加密等，進一步加強對個人信息的保護，降低隱私泄露的安全風(fēng)險。本文首先分析了聯(lián)邦學(xué)習(xí)是基于設(shè)計隱私思想的分布式協(xié)作模型，然后對聯(lián)邦學(xué)習(xí)框架在個人信息保護原則下的表現(xiàn)進行評價并提出建議，最后探討了聯(lián)邦學(xué)習(xí)如何促進數(shù)據(jù)合規(guī)并指出依然存在的合規(guī)風(fēng)險。
本文詳細內(nèi)容請下載：http://m.ihrv.cn/resource/share/2000005369
作者信息：
孫綺雯
(清華大學(xué)法學(xué)院，北京100084)
此內(nèi)容為AET網(wǎng)站原創(chuàng)，未經(jīng)授權(quán)禁止轉(zhuǎn)載。