边缘计算下的轻量级联邦学习隐私保护方案

Lightweight Federated Learning Privacy Protection Scheme under Edge Computing

随着物联网和大数据技术的高速发展,以传统云计算模式为代表的集中式学习效率低下,且易受到单点攻击、共谋攻击、中间人攻击等一系列攻击手段,造成数据安全的隐患。边缘计算模式使得分布式联邦学习成为了可能,然而,联邦学习虽然能够保证数据在本地的安全和隐私,但是也面临众多安全威胁,如梯度泄露攻击,此外,效率问题也是联邦学习的痛点所在。为了保障边缘计算场景下的模型训练安全,提出了一种边缘计算下的轻量级联邦学习隐私保护方案(Lightweight Federated Learning Privacy Protection Scheme Under Edge Computing,LFLPP)。首先,提出一种云-边-端分层的联邦学习框架;其次,对不同层进行隐私保护;最后,提出一种周期性更新策略,极大地提高了收敛速度。使用乳腺癌肿瘤数据集和CIFAR10数据集在LR模型和Resnet18残差模型上进行训练和测试,同时使用CIFAR10数据集与FedAvg和PPFLEC(Privacy-Preserving Federated Learning for Internet of Medical Things under Edge Computing)两种方案进行对比实验,得出准确率和训练效率的差距,并进行准确率、效率以及安全性等方面的分析,该方案在CIFAR-10数据集上能达到84.63%的准确率。

With the rapid development of the Internet of Things and big data technology,centralized learning represented by the traditional cloud computing model is inefficient and vulnerable to a series of attacks such as single point attack,collusion attack,man in the middle attack,resulting in hidden dangers of data security.The edge computing model makes distributed federated learning possible.However,although federated learning can ensure the security and privacy of data locally,it also faces many security threats,such as gradient disclosure attacks.In addition,the efficiency is also the pain point of federated learning.In order to ensure the security of model training in the edge computing scenario,a lightweight federated learning privacy protection scheme under edge computing(LFLPP)is proposed.Firstly,a cloud-edge-end layered federated learning framework is proposed.Secondly,privacy protection for different layers.Finally,a periodic updating strategy is proposed,which greatly improves the convergence speed.The breast cancer tumor data set and CIFAR10 data set were used for training and testing on LR model and Resnet18 residual model.At the same time,CIFAR10 data set was used to conduct comparative experiments with FedAvg and PPFLEC(Privacy Preserving Federated Learning for Internet of Medical Things under Edge Computing),to find out the gap between accuracy and training efficiency,and to conduct accuracy analysis,efficiency analysis and security analysis,This scheme can achieve 84.63%accuracy on CIFAR-10 dataset.