Secure federated learning at the edge

确保边缘联合学习的安全

• 批准号: 568539-2021
• 负责人: Agarwal, Anjali
• 金额: $ 2.19万
• 依托单位: Concordia University
• 依托单位国家: 加拿大
• 项目类别: Alliance Grants
• 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=723193
• 关键词: Secure; federated; learning; edge

Federated Learning (FL) is emerging as one of the fundamental Machine Learning (ML) models that supports distributed model training with privacy-preserving characteristics. Ever since its inception in 2017 by Google AI, it is being utilized across different verticals such as e-healthcare, smart grids, industrial sector, insurance sector, autonomous vehicles, FinTech, etc. Some of the characteristic advantages of FL against the traditional ML models include data privacy and diversity, real-time data analysis even during low/no Internet connectivity, hardware efficiency, etc. Despite these advantages, FL is founded vulnerable to different cyberattacks such as data and model poisoning, inferences, backdoors, malicious server, communication bottlenecks, etc.Thus, as part of this research project, we propose to develop an efficient, secure, and privacy-aware FL framework that can be leveraged across different application domains using the Edge computing paradigm. The research will employee the advantages of differential privacy, homomorphic encryption, and blockchain; and the designed framework will be validated on real-time and benchmark datasets for enhanced efficacy. This research will allow the Canada-based company Cistech Ltd to develop more robust, secure, and privacy-preserving FL-based solutions for its client, paying special attention to the on-going outbreak of different attack vectors on the ML models with increasing complexity and sophistication. Thus, the company will be able to ensure the deployment of secure FL-based solutions to it clients. The designed solution during this research will be transferred to industry and can be deployed in their upcoming projects across different domains.

联邦学习 (FL) 正在成为基本机器学习 (ML) 模型之一，支持具有隐私保护特性的分布式模型训练。自 Google AI 于 2017 年推出以来，它已被应用于不同的垂直领域，例如电子医疗、智能电网、工业部门、保险部门、自动驾驶汽车、金融科技等。FL 相对于传统 ML 的一些特征优势模型包括数据隐私和多样性、即使在低/无互联网连接情况下的实时数据分析、硬件效率等。尽管有这些优势，FL 仍然容易受到不同的网络攻击，例如数据和模型中毒、推理、后门、恶意服务器、通信瓶颈等。因此，作为本研究项目的一部分，我们建议开发一种高效、安全且具有隐私意识的 FL 框架，该框架可以使用边缘计算范例在不同的应用领域中使用。该研究将利用差分隐私、同态加密和区块链的优势；设计的框架将在实时和基准数据集上进行验证，以提高效率。这项研究将使加拿大公司 Cistech Ltd 能够为其客户开发更强大、更安全和保护隐私的基于 FL 的解决方案，并特别关注 ML 模型上不断爆发的不同攻击向量，且复杂性不断增加和复杂程度。因此，该公司将能够确保为其客户部署安全的基于 FL 的解决方案。本研究期间设计的解决方案将转移到行业，并可以部署在不同领域的即将推出的项目中。