面向隐私敏感物联网数据的联邦学习双向通信压缩

doi:10.11772/j.issn.1001-9081.2025040436

《计算机应用》唯一官方网站 ›› 2026, Vol. 46 ›› Issue (3): 887-898.DOI: 10.11772/j.issn.1001-9081.2025040436

面向隐私敏感物联网数据的联邦学习双向通信压缩

王磊¹, 周文轩², 贾柠晖², 屈志昊²()

^1.国网江苏省电力有限公司信息通信分公司，南京 210024
^2.水利部水利大数据重点实验室（河海大学），南京 211100

收稿日期:2025-04-22 修回日期:2025-07-09 接受日期:2025-07-11 发布日期:2025-07-22 出版日期:2026-03-10
通讯作者: 屈志昊
作者简介:王磊（1978—），男，江苏扬州人，高级工程师，硕士，主要研究方向：电力物联网、边缘智能
周文轩（1999—），男（回族），江苏扬州人，博士研究生，CCF会员，主要研究方向：联邦学习、分割学习
贾柠晖（1998—），男，江苏无锡人，博士研究生，CCF会员，主要研究方向：联邦学习中的通信优化、联邦学习中的无梯度优化
基金资助:
国网江苏省电力有限公司科技项目(J2023077)

Federated learning with two-pass communication compression for privacy-sensitive IoT data

Lei WANG¹, Wenxuan ZHOU², Ninghui JIA², Zhihao QU²()

^1.Information and Telecommunication Branch，State Grid Jiangsu Electric Power Company Limited，Nanjing Jiangsu 210024，China
^2.Key Laboratory of Water Big Data Technology of Ministry of Water Resources （Hohai University），Nanjing Jiangsu 211100，China

Received:2025-04-22 Revised:2025-07-09 Accepted:2025-07-11 Online:2025-07-22 Published:2026-03-10
Contact: Zhihao QU
About author:WANG Lei， born in 1978， M. S.， senior engineer. His research interests include electric power internet of things， edge intelligence.
ZHOU Wenxuan， born in 1999， Ph. D. candidate. His research interests include federated learning， split learning.
JIA Ninghui， born in 1998， Ph. D. candidate. His research interests include communication optimization in federated learning， gradient-free optimization in federated learning.
Supported by:
Science and Technology Project of State Grid Jiangsu Electric Power Company Limited(J2023077)

摘要/Abstract

摘要：

针对隐私敏感物联网（IoT）场景中联邦学习面临巨大通信开销和梯度反演隐私泄露风险问题，提出一种名为QPR（Quantization and Pull Reduction）的两阶段联邦学习通信压缩框架。首先，将训练节点通过梯度量化技术压缩本地梯度后上传至服务器，以降低梯度传输的开销；其次，引入基于概率阈值的延迟模型下载机制（lazy pulling）以降低模型的同步频率，训练节点以预设概率同步全局模型，而其余迭代中复用本地历史模型；最后，通过严格的理论分析确保QPR在收敛速度上达到与标准无通信压缩联邦学习——联邦平均（FedAvg）算法相同的渐进阶次，且具备随训练节点数增多而线性加速的特性，从而保证系统的可延展性。实验结果表明，QPR在多个基准数据集和机器学习模型上均能显著提升通信效率。以ResNet18模型在CIFAR-10数据集上的训练任务为例，QPR在不损失模型精度的前提下，与无压缩的FedAvg算法相比，最高实现了8.27的通信加速比。

关键词: 联邦学习, 梯度量化, 两阶段压缩, 物联网, 隐私保护

Abstract:

To address the significant communication overhead and gradient inversion privacy leakage risks of federated learning in Internet of Things （IoT） scenarios， a two-pass communication compression framework named QPR （Quantization and Pull Reduction） was proposed. Firstly， through gradient quantization， training nodes were utilized to compress local gradients before uploading them to the server， thereby reducing gradient transmission overhead. Secondly， a probability threshold-based delayed model download mechanism （lazy pulling） was introduced to reduce model synchronization frequency， and the global model was synchronized by training nodes. For other iterations， locally historical models were reused. Finally， rigorous theoretical analysis was performed to confirm that QPR achieves the same progressive order as the standard federated learning algorithm without communication compression — Federated Average （FedAvg） and possesses the linear speedup property with increasing number of nodes， thereby ensuring system scalability. Experimental results demonstrate that QPR enhances communication efficiency on multiple benchmark datasets and machine learning models significantly. Taking the ResNet18 model training task on the CIFAR-10 dataset as an example， QPR achieves communication speedup ratio up to 8.27 compared to uncompressed FedAvg， without any loss in model accuracy.

Key words: Federated Learning (FL), gradient quantization, two-pass compression, Internet of Things (IoT), privacy protection

中图分类号:

TP391

王磊, 周文轩, 贾柠晖, 屈志昊. 面向隐私敏感物联网数据的联邦学习双向通信压缩[J]. 计算机应用, 2026, 46(3): 887-898.

Lei WANG, Wenxuan ZHOU, Ninghui JIA, Zhihao QU. Federated learning with two-pass communication compression for privacy-sensitive IoT data[J]. Journal of Computer Applications, 2026, 46(3): 887-898.

图/表 11

图1 联邦学习的结构

Fig. 1 Structure of federated learning

图2 QPR在4个IoT设备的联邦学习迭代t中的示例

Fig. 2 Example of QPR in iteration t of 4-IoT-device federated learning

图3 不同量化比特数下的机器学习任务结果

Fig. 3 Results of machine learning tasks under different number of quantization bits

图4 不同拉取比率下的机器学习任务性能对比

Fig. 4 Results of machine learning tasks under different pulling ratios

图5 不同规模IoT设备集群的机器学习多任务性能对比

Fig. 5 Comparison of machine learning multi-task performance onclusters with different number of IoT devices

图6 不同压缩方案的收敛性与泛化性对比

Fig. 6 Comparison of convergence and generalization of different compression schemes

表1 不同压缩算法达到NSGD基准精度时的带宽节省倍数对比

Tab. 1 Comparison of bandwidth savings when different compression algorithms achieving NSGD benchmark accuracy

精度/%	QPR	DQSGD	QSGD	PR
90	7.30×	4.97×	1.68×	1.55×
80	6.99×	6.04×	1.89×	1.49×
70	7.37×	6.72×	1.93×	1.64×
60	8.27×	7.66×	2.10×	2.05×

表2 不同压缩算法达到NSGD基准精度时的总通信时延

Tab. 2 Total communication delays when different compression algorithms achieving NSGD benchmark accuracy

精度/%	总通信时延/s
精度/%	NSGD	QPR	DQSGD	QSGD	PR
90	24 481.5	3 355.4	4 929.8	14 612.2	15 807.7
80	12 280.1	1 756.4	2 031.9	6 509.1	8 236.0
70	6 415.6	870.8	954.5	3 320.9	3 918.7
60	3 581.7	433.0	467.4	1 704.8	1 749.0

图7 QPR采用不同学习率的收敛性与泛化性对比

Fig. 7 Comparison of convergence and generalization of QPR using different learning rates

图8 不同异构程度下多种压缩算法的精度对比

Fig. 8 Comparison of accuracy of multiple compression algorithms under different heterogeneity levels

图9 QPR算法与缓解异构性算法的兼容性

Fig. 9 Compatibility between QPR algorithm and heterogeneity mitigation algorithms

参考文献 39

[1]	KOLOSKOVA A， STICH S U， JAGGI M. Sharper convergence guarantees for asynchronous SGD for distributed and federated learning ［C］// Proceedings of the 36th International Conference on Neural Information Processing Systems. Red Hook： Curran Associates Inc.， 2022： 17202-17215.
[2]	葛丽娜，王明禹，田蕾. 联邦学习的高效性研究综述［J］. 计算机应用， 2025， 45（8）： 2387-2398.
	GE L N， WANG M Y， TIAN L. Review of research on efficiency of federated learning ［J］. Journal of Computer Applications， 2025， 45（8）： 2387-2398.
[3]	ALISTARH D， GRUBIC D， LI J Z， et al. QSGD： communication-efficient SGD via gradient quantization and encoding ［C］// Proceedings of the 31st International Conference on Neural Information Processing Systems. Red Hook： Curran Associates Inc.， 2017： 1707-1718.
[4]	HÖNIG R， ZHAO Y， MULLINS R. DAdaQuant： doubly-adaptive quantization for communication-efficient federated learning ［C］// Proceedings of the 39th International Conference on Machine Learning. New York： JMLR.org， 2022： 8852-8866.
[5]	ZHANG Z， WANG C. MIPD： an adaptive gradient sparsification framework for distributed DNNs training ［J］. IEEE Transactions on Parallel and Distributed Systems， 2022， 33（11）： 3053-3066.
[6]	ZHANG J， SIMEONE O. LAGC： lazily aggregated gradient coding for straggler-tolerant and communication-efficient distributed learning ［J］. IEEE Transactions on Neural Networks and Learning Systems， 2021， 32（3）： 962-974.
[7]	邱鑫源，叶泽聪，崔翛龙，等. 联邦学习通信开销研究综述［J］. 计算机应用， 2022， 42（2）： 333-342.
	QIU X Y， YE Z C， CUI X L， et al. Survey on communication overhead in federated learning［J］. Journal of Computer Applications， 2022， 42（2）： 333-342.
[8]	YU Y， WU J， HUANG L. Double quantization for communication-efficient distributed optimization ［C］// Proceedings of the 33rd International Conference on Neural Information Processing Systems. Red Hook： Curran Associates Inc.， 2019： 4438-4449.
[9]	TANG H， LIAN X， YU C， et al. DoubleSqueeze： parallel stochastic gradient descent with double-pass error-compensated compression ［C］// Proceedings of the 36th International Conference on Machine Learning. New York： JMLR.org， 2019： 6155-6165.
[10]	张瑞麟，杜晋华，尹浩. 跨设备联邦学习中的客户端选择算法［J］. 软件学报， 2024， 35（12）： 5725-5740.
	ZHANG R L， DU J H， YIN H. Client selection algorithm in cross-device federated learning ［J］. Journal of Software， 2024， 35（12）： 5725-5740.
[11]	QIAO A， ARAGAM B， ZHANG B， et al. Fault tolerance in iterative-convergent machine learning ［C］// Proceedings of the 36th International Conference on Machine Learning. New York： JMLR.org， 2019： 5220-5230.
[12]	SAFARYAN M， SHULGIN E， RICHTÁRIK P. Uncertainty principle for communication compression in distributed and federated learning and the search for an optimal compressor ［J］. Information and Inference： A Journal of the IMA， 2022， 11（2）： 557-580.
[13]	WANG M， BODONHELYI A， BOZKIR E， et al. TurboSVM-FL： boosting federated learning through SVM aggregation for lazy clients［C］// Proceedings of the 38th AAAI Conference on Artificial Intelligence. Palo Alto： AAAI Press， 2024： 15546-15554.
[14]	CRAWSHAW M， LIU M. Federated learning under periodic client participation and heterogeneous data： a new communication-efficient algorithm and analysis ［C］// Proceedings of the 38th International Conference on Neural Information Processing Systems. Red Hook： Curran Associates Inc.， 2024： 8240-8299.
[15]	ZHANG W， ZHOU T， LU Q， et al. FedSL： a communication-efficient federated learning with split layer aggregation ［J］. IEEE Internet of Things Journal， 2024， 11（9）： 15587-15601.
[16]	QU Z， JIA N， YE B， et al. FedQClip： accelerating federated learning via quantized clipped SGD ［J］. IEEE Transactions on Computers， 2025， 74（2）： 717-730.
[17]	SEIDE F， FU H， DROPPO J， et al. 1-bit stochastic gradient descent and its application to data-parallel distributed training of speech DNNs ［C］// Proceedings of the INTERSPEECH 2014. ［S.l.］： International Speech Communication Association， 2014： 1058-1062.
[18]	WEN W， XU C， YAN F， et al. TernGrad： ternary gradients to reduce communication in distributed deep learning ［C］// Proceedings of the 31st International Conference on Neural Information Processing Systems. Red Hook： Curran Associates Inc.， 2017： 1508-1518.
[19]	ZHOU Q， GUO S， LIU Y， et al. Hierarchical channel-spatial encoding for communication-efficient collaborative learning ［C］// Proceedings of the 36th International Conference on Neural Information Processing Systems. Red Hook： Curran Associates Inc.， 2022： 5788-5801.
[20]	ZAKERINIA H， TALAEI S， NADIRADZE G， et al. Communication-efficient federated learning with data and client heterogeneity ［C］// Proceedings of the 27th International Conference on Artificial Intelligence and Statistics. New York： JMLR.org， 2024： 3448-3456.
[21]	CHEN Y， VIKALO H， WANG C. Fed-QSSL： a framework for personalized federated learning under bitwidth and data heterogeneity ［C］// Proceedings of the 38th AAAI Conference on Artificial Intelligence. Palo Alto： AAAI Press， 2024： 11443-11452.
[22]	ZHANG P， XU L， MEI L， et al. Sketch-based adaptive communication optimization in federated learning ［J］. IEEE Transactions on Computers， 2025， 74（1）： 170-184.
[23]	ZHOU W， QU Z， LYU S H， et al. Mask-encoded sparsification： mitigating biased gradients in communication-efficient split learning［C］// Proceedings of the 27th European Conference on Artificial Intelligence. Amsterdam： IOS Press， 2024： 2806-2813.
[24]	ALBELAIHI R， ALASANDAGUTTI A， YU L， et al. Deep-reinforcement-learning-assisted client selection in nonorthogonal-multiple-access-based federated learning ［J］. IEEE Internet of Things Journal， 2023， 10（17）： 15515-15525.
[25]	KIM G， KIM J， HAN B. Communication-efficient federated learning with accelerated client gradient ［C］// Proceedings of the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2024： 12385-12394.
[26]	LI J， LIU Y， WANG W. FedNS： a fast sketching Newton-type algorithm for federated learning ［C］// Proceedings of the 38th AAAI Conference on Artificial Intelligence. Palo Alto： AAAI Press， 2024： 13509-13517.
[27]	GONG X， LI S， BAO Y， et al. Federated learning via input-output collaborative distillation ［C］// Proceedings of the 38th AAAI Conference on Artificial Intelligence. Palo Alto： AAAI Press， 2024： 22058-22066.
[28]	ELBAKARY A， ISSAID C B， SHEHAB M， et al. Fed-Sophia： a communication-efficient second-order federated learning algorithm［C］// Proceedings of the 2024 IEEE International Conference on Communications. Piscataway： IEEE， 2024： 950-955.
[29]	CHEN T， GIANNAKIS G， SUN T， et al. LAG： lazily aggregated gradient for communication-efficient distributed learning ［C］// Proceedings of the 32nd International Conference on Neural Information Processing Systems. Red Hook： Curran Associates Inc.， 2018： 5055-5065.
[30]	KIM D Y， HAN D J， SEO J， et al. Achieving lossless gradient sparsification via mapping to alternative space in federated learning［C］// Proceedings of the 41st International Conference on International Conference on Machine Learning. New York： JMLR.org， 2024： 23867-23900.
[31]	SHOKRI R， SHMATIKOV V. Privacy-preserving deep learning［C］// Proceedings of the 22nd ACM SIGSAC Conference on Computer and Communications Security. New York： ACM， 2015： 1310-1321.
[32]	TANG Z， HUANG J， YAN R， et al. Bandwidth-aware and overlap-weighted compression for communication-efficient federated learning ［C］// Proceedings of the 53rd International Conference on Parallel Processing. New York： ACM， 2024： 866-875.
[33]	XIE W， LI H， MA J， et al. JointSQ： Joint sparsification-quantization for distributed learning ［C］// Proceedings of the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2024： 5778-5787.
[34]	KRIZHEVSKY A. Learning multiple layers of features from tiny images ［R/OL］. ［2025-03-06］..
[35]	苏瑞，朱亚丽，朱晓荣. 面向6G网络的联邦学习高效训练优化方法［J/OL］. 物联网学报［2025-03-08］..
	SU R， ZHU Y L， ZHU X R. An efficient training optimization method for federated learning in 6G networks ［J/OL］. Chinese Journal of Internet of Things ［2025-03-08］..
[36]	ZHENG S， SHEN C， CHEN X. Design and analysis of uplink and downlink communications for federated learning［J］. IEEE Journal on Selected Areas in Communications， 2021， 39（7）： 2150-2167.
[37]	HOSMER D W， HOSMER T， LE CESSIE S， et al. A comparison of goodness-of-fit tests for the logistic regression model ［J］. Statistics in Medicine， 1997， 16（9）： 965-980.
[38]	KRIZHEVSKY A， SUTSKEVER I， HINTON G E. ImageNet classification with deep convolutional neural networks ［C］// Proceedings of the 25th International Conference on Neural Information Processing Systems — Volume 1. Red Hook： Curran Associates Inc.， 2012： 1097-1105.
[39]	KARIMIREDDY S P， REBJOCK Q， STICH S U， et al. Error feedback fixes SignSGD and other gradient compression schemes［C］// Proceedings of the 36th International Conference on Machine Learning. New York： JMLR.org， 2019： 3252-3261.

面向隐私敏感物联网数据的联邦学习双向通信压缩

Federated learning with two-pass communication compression for privacy-sensitive IoT data

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 11

参考文献 39

相关文章 15

编辑推荐

Metrics

[1]	马凯光, 陈学斌, 菅银龙, 王柳, 高远. 基于混合序列模型与联邦类平衡算法的网络入侵检测[J]. 《计算机应用》唯一官方网站, 2026, 46(3): 857-866.
[2]	平欢, 夏战国, 刘思诚, 刘奇翰, 李春磊. 基于多层联邦学习的终端数据隐私保护方案[J]. 《计算机应用》唯一官方网站, 2026, 46(3): 830-838.
[3]	郗恩康, 范菁, 金亚东, 董华, 俞浩, 孙伊航. 联邦学习在隐私安全领域面临的威胁综述[J]. 《计算机应用》唯一官方网站, 2026, 46(3): 798-808.
[4]	何丽丽, 管新如, 张磊, 蒋胜, 蒋澄杰. 车联网位置隐私保护的全景与未来[J]. 《计算机应用》唯一官方网站, 2026, 46(3): 809-820.
[5]	钟琪, 张淑芬, 张镇博, 菅银龙, 景忠瑞. 面向联邦学习的投毒攻击检测与防御机制[J]. 《计算机应用》唯一官方网站, 2026, 46(2): 445-457.
[6]	何金栋, 及宇轩, 陈天赐, 许恒铭, 耿技, 曹明生, 梁员宁. 基于知识图谱和大模型的非智能传感器的实体发现方法[J]. 《计算机应用》唯一官方网站, 2026, 46(2): 354-360.
[7]	张珂嘉, 方志军, 周南润, 史志才. 基于模型预分配与自蒸馏的个性化联邦学习方法[J]. 《计算机应用》唯一官方网站, 2026, 46(1): 10-20.
[8]	樊娜, 罗闯, 张泽晖, 张梦瑶, 穆鼎. 基于改进生成对抗网络的车辆轨迹语义隐私保护机制[J]. 《计算机应用》唯一官方网站, 2026, 46(1): 169-180.
[9]	俞浩, 范菁, 孙伊航, 金亚东, 郗恩康, 董华. 边缘异构下的联邦分割学习优化方法[J]. 《计算机应用》唯一官方网站, 2026, 46(1): 33-42.
[10]	菅银龙, 陈学斌, 景忠瑞, 钟琪, 张镇博. 联邦学习中基于条件生成对抗网络的数据增强方案[J]. 《计算机应用》唯一官方网站, 2026, 46(1): 21-32.
[11]	翟社平, 朱鹏举, 杨锐, 刘佳一腾. 基于区块链的物联网身份管理系统[J]. 《计算机应用》唯一官方网站, 2025, 45(9): 2873-2881.
[12]	张博瀚, 吕乐, 荆军昌, 刘栋. 基于遗传算法的属性网络社区隐藏方法[J]. 《计算机应用》唯一官方网站, 2025, 45(9): 2817-2826.
[13]	俞浩, 范菁, 孙伊航, 董华, 郗恩康. 联邦学习统计异质性综述[J]. 《计算机应用》唯一官方网站, 2025, 45(9): 2737-2746.
[14]	苏锦涛, 葛丽娜, 肖礼广, 邹经, 王哲. 联邦学习中针对后门攻击的检测与防御方案[J]. 《计算机应用》唯一官方网站, 2025, 45(8): 2399-2408.
[15]	葛丽娜, 王明禹, 田蕾. 联邦学习的高效性研究综述[J]. 《计算机应用》唯一官方网站, 2025, 45(8): 2387-2398.