联邦算力物联网(IoT,Internet of things)旨在通过联邦学习深度融合算力与物联网资源,从而实现对泛在离散部署的海量物联网数据和异构资源的高效利用.为了应对联邦算力物联网中模型反演和梯度泄露等新兴隐私攻击威胁,学术界和产业界对差分隐私(DP,differential privacy)这一高效的隐私保护技术进行了广泛研究和应用.然而,现有差分隐私技术在设定隐私预算时,未考虑本地算力节点的数据特征和隐私预算分配公平性的问题,造成了严重的模型精度损失.因此,提出了一种面向联邦算力物联网的隐私预算自适应优化方案——基于克拉美罗下界差分隐私的联邦学习(FedCDP,federated learning based on Cramér-Rao lower bound differential privacy).首先,基于克拉美罗下界理论分析边缘算力节点的隐私预算估计值,实现自适应隐私预算规划;其次,通过计算边缘算力节点的上传模型与算力聚合服务器的聚合模型之间的相似度和隐私预算占比,分析得到每个节点的全局贡献度,进一步联合隐私预算估计值公平实时地优化隐私预算设定.理论分析证明了该方案可确保本地模型严格遵守ε-差分隐私,并保证全局模型收敛.基于多个公开数据集上的实验结果表明,在满足相同隐私保护需求的前提下,该方案将全局模型精确度最多提升了10.19%.
A privacy budget adaptive optimization scheme for federated computing power Internet of things
Federated computing power Internet of things(IoT)is designed to deeply integrate computing power with IoT resources,facilitating the efficient utilization of vast and ubiquitously dispersed IoT data and heterogeneous resources through federated learning.Faced with the threats of emerging privacy attacks,e.g.,model inversion attacks and gradient leakage attacks,the academic and industrial communities have widely investigated and applied differential privacy(DP)as an effective privacy protection technique.However,two severe challenges have not been taken into account in the exist-ing DP budget settings,i.e.,data heterogeneity issue of local computing power nodes and the fairness of privacy budget al-location,which lead to a significant loss in model accuracy.Therefore,an adaptive optimization scheme for privacy bud-get was proposed in federated computing power IoT,which was called federated learning based on Cramér-Rao lower bound differential privacy(FedCDP).In specific,to adaptively adjust privacy budgets,the privacy budget estimates for edge computing power nodes based on the Cramér-Rao lower bound theory were analyzed.Furthermore,by assessing the simi-larity between the local model and the aggregated model,as well as their respective privacy budget proportions,the global contribution of each node was determined,which was used to fairly,also in real time,optimize and adjust the privacy bud-get settings in conjunction with the estimated privacy budget.Through rigorous theoretical analysis,FedCDP achieves ε-DP for local models,and ensures the convergence of the global model.Experimental results on multiple public datasets show that the proposed scheme improves the accuracy of the global model by up to 10.19%under the premise of satisfy-ing the same privacy protection requirements.
federated computing power IoTdifferential privacyprivacy budgetadaptive optimization
国家科技期刊平台
NSTL
万方数据
