Privacy-preserving machine learning through secure management of data's lifecycle in distributed systems: REMINDER

人工智能(AI)变得无处不在，并引领着一场技术范式的转变。联合国2030年可持续发展目标(SDGs)中设定的一些主要目标将需要通过负责任地使用人工智能技术来解决，以将数据转化为真正的知识，造福我们的社会。这一趋势是通过基于分布式系统与互联网基础设施的集成(主要基于物联网(IoT)技术和5G/6G基础设施的部署)的超级连接程度不断提高来推动的。这些系统的整合将在我们周围的环境中实现新的基于数据的服务，例如在可持续城市和社区或先进的电子健康服务的背景下。为了有效和高效地提供这些服务，一个关键方面是在数据的整个生命周期中管理安全和隐私，以确保服务基于合法系统提供的可信信息。在这个方向上，该项目(提醒)将设计一种分散和安全的方法，用于获取和处理分布式系统产生的数据。特别是，Remder将为使用联合学习(FL)的应用程序设计和实施基于边缘的体系结构，该体系结构将可供资源受限的终端节点访问。与当前的大多数部署不同，该架构将支持协作模型创建，而无需共享数据本身。该体系结构将通过为FL体系结构中的训练过程设计节点选择方法来考虑分散和分布式系统的高度动态化，同时考虑终端系统的特征(例如，设备状态或电池水平)以及它们在其生命周期中的演变。此外，提醒将解决与使用去中心化机器学习(ML)方法相关的一些主要安全和隐私挑战，如FL。在这个方向上，该项目将分析加密技术的使用，如差分隐私(DP)和安全多方计算(SMPC)，以便在ML技术提供的有效性和所保证的隐私水平之间取得适当的平衡。在休息、传输和处理过程中将考虑数据隐私。拟议的解决方案将是预防性和反应性的。他们还将确保隐私保护方面符合现有的数据保护法规，如数据生命周期的GDPR。提醒还将通过设计和实施身份验证协议来解决FL环境中的一些主要安全攻击，以确保只有合法系统才能参与ML模型的协作创建过程。此外，提醒将通过围绕电子健康和智能建筑的两个主要用例来论证拟议研究的可行性。

The Artificial Intelligence (AI) becomes ubiquitous and leading a technological paradigm shift. Some of the main objectives set out in the United Nations' Sustainable Development Goals (SDGs) for 2030 will require to be addressed through the responsible use of AI techniques to transform data into real knowledge for the benefit of our society. This trend is being driven through an increasing degree of hyperconnectivity based on the integration of distributed systems into the Internet infrastructure mainly based on the deployment of Internet of Things (IoT) technologies as well as 5G/6G infrastructures. The integration of such systems will enable new data-based services in our surrounding environment, e.g., in the context of sustainable cities and communities or advanced eHealth services. To provide these services effectively and efficiently, a key aspect is the management of security and privacy throughout the data's lifecycle in a way that ensures the services are based on trustworthy information provided by legitimate systems. In this direction, this project (REMINDER) will design a decentralized and secure approach for the access and processing of data produced by distributed systems. In particular, REMINDER will design and implement an edge-based architecture for applications using Federated Learning (FL) that will be accessible to resource-constrained end nodes. Unlike most current deployments, the architecture will enable a collaborative model creation without the need to share the data itself. This architecture will consider the high degree of dynamism of decentralized and distributed systems by designing a node selection approach for the training process in the FL architecture while considering end systems' features (e.g., device status or battery level), as well as their evolution during their life cycle. Additionally, REMINDER will address some of the major security and privacy challenges associated with the use of decentralized Machine Learning (ML) approaches, such as FL. In this direction, the project will analyze the use of cryptographic techniques, such as Differential Privacy (DP) and Secure Multi- Party Computation (SMPC) for the sake of reaching the right balance between the effectiveness provided by ML techniques and the level of privacy being guaranteed. Data privacy will be considered in rest, transit, and while processing. The proposed solutions will be preventive and reactive. They will also ensure the privacy preserving aspects are being compliant with existing data protection regulations, such as the GDPR over the data lifecycle. REMINDER will also address some of the major security attacks in FL environments by designing and implementing an authentication protocol to ensure that only legitimate systems are able to take part in the collaborative creation process of ML models. Moreover, REMINDER will demonstrate the feasibility of the proposed research through two main use cases around eHealth and smart buildings.