Efficient Secure Distributive Machine Learning over Mobile Network Systems

移动网络系统上的高效安全分布式机器学习

• 批准号: RGPIN-2019-05348
• 负责人: Gong, Guang
• 金额: $ 3.35万
• 依托单位: University of Waterloo
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2021
• 资助国家: 加拿大
• 起止时间: 2021-01-01 至 2022-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=738211
• 关键词: Efficient; Secure; Distributive; Machine; Learning

In recent years, the use of smartphones, wearables, and tablet applications is growing, and extensively increasing. Mobile applications use a large amount of sensitive data such as users' information and location, health care data, banking information, and bioinformatics. The Internet--of--Things (IoT) is another resource of massive big data, and future 5G cellular systems aim to support its applications. Machine learning (ML) has risen as the power of decision-making processes in many practices, e.g., disease control by trend prediction in healthcare, predictions for financial markets in finance, intelligent consumer targeting for economic growth, traffic patterns and energy usage in infrastructure, and spam detection, traffic analysis and intrusion detection in network defence systems, to name a few. However, a server handling these requests needs to access a large amount of sensitive data from devices and users, which puts the security and privacy of users' data at immense risk. Cloud service providers such as Google, Amazon, Microsoft, and IBM provide machine--learning--as--a--service (MLaaS) tools, such as data visualization, APIs, and others as part of their service. The proposed research considers a machine learning model where sensitive data remains on the mobile devices (smart phones, sensors, automotive on-board units, IoT devices, etc.) that process  some ML algorithms locally. In this scenario, training data is distributed to mobile devices, and the server/cloud learns a shared model using locally computed updates. We term this as distributive machine learning over mobile networks (DMLMN) model.  The problem that we are facing is providing the security and privacy of data against the server/cloud, while maintaining the server/cloud's ability to perform the necessary computations in ML algorithms. The goal of the proposed research is therefore to investigate a new paradigm for designing algorithms and protocols with efficient computation, communication, and cyber-attack resiliency for secure DMLMN. We will adopt a new paradigm of mask--then--encrypt with noise cancellation and evaluate the performance and tradeoffs of algorithms and protocols under active malicious attacks (e.g., mobile device or server may be malicious, attacker may craft the training data which results in a bias model, how to detect them, etc.). We will attempt to achieve this goal through combinations of cryptography, coding theory, and physical layer security. The outcome will provide new efficient and practical algorithms and protocols for distributive secure mobile learning. The research will benefit sectors such as financial, bioinformatics, health and medical, Internet and cloud data industrials, and data market. HQP trained in this unique environment will enrich the talent pool of engineers and researchers in Canada with the expertise to create innovative solutions to the evolving challenges of developing secure  DMLMN.

近年来，智能手机、可穿戴设备和平板电脑应用的使用正在增长，并且广泛增加。移动的应用程序使用大量敏感数据，例如用户的信息和位置、医疗保健数据、银行信息和生物信息学。物联网（IoT）是另一个海量大数据资源，未来的5G蜂窝系统旨在支持其应用。机器学习（ML）在许多实践中已经成为决策过程的力量，例如，通过医疗保健中的趋势预测进行疾病控制、金融中的金融市场预测、针对经济增长的智能消费者定位、基础设施中的流量模式和能源使用、以及网络防御系统中的垃圾邮件检测、流量分析和入侵检测，仅举几例。然而，处理这些请求的服务器需要访问来自设备和用户的大量敏感数据，这使得用户数据的安全性和隐私性面临巨大风险。Google、Amazon、Microsoft和IBM等云服务提供商提供机器学习即服务（MLaaS）工具，如数据可视化、API等，作为其服务的一部分。 拟议的研究考虑了一种机器学习模型，其中敏感数据保留在移动的设备（智能手机，传感器，车载单元，物联网设备等）上。在本地处理一些ML算法。在这种情况下，训练数据被分发到移动的设备，并且服务器/云使用本地计算的更新来学习共享模型。我们称之为基于移动的网络的分布式机器学习（DMLMN）模型。我们面临的问题是针对服务器/云提供数据的安全性和隐私性，同时保持服务器/云在ML算法中执行必要计算的能力。因此，所提出的研究的目标是研究一种新的范式，用于设计具有有效计算，通信和网络攻击弹性的算法和协议，以实现安全的DMLMN。我们将采用一种新的掩码-然后-加密与噪声消除的范例，并评估算法和协议在主动恶意攻击（例如，移动终端或服务器可能是恶意的，攻击者可能精心制作导致偏差模型的训练数据，如何检测它们等）。我们将尝试通过密码学、编码理论和物理层安全性的组合来实现这一目标。研究结果将为分布式安全移动的学习提供新的高效实用的算法和协议。该研究将使金融，生物信息学，健康和医疗，互联网和云数据产业以及数据市场等部门受益。在这种独特的环境中接受培训的HQP将丰富加拿大工程师和研究人员的人才库，他们将拥有专业知识，为开发安全DMLMN的不断变化的挑战创造创新的解决方案。