CN2Conv：面向物联网设备的强鲁棒CNN设计方法

打开文本图片集

关键词：卷积神经网络；云辅助训练；少参数模型；鲁棒性

中图分类号：TP183 文献标志码：A 文章编号：1001-3695（2025）07-031-2154-07

doi： 10.19734/j . issn.1001-3695.2024.12.0500

Abstract：Theuseofcloud-assisted training forCNN with fewparameterscanenable theirdeploymentonresource-constrained IoTdevices.However，existing models with fewparameters suferfrom insuficientabilitytoextractcomplexdatafeatures and poor robustnessThis article proposed a CNN design method that was adaptable to complexdata and hadstrong robustness， calledthecombinednon-linearityconvolution kernelgeneration（CN2Conv）.Firstly，itrandomlyselectedsomeconvolution kernels fromtheconvolutionallayersoftheCNNmodelasseedconvolution kernels，andusedmultiplegeneration functions to performonlinear transformationsontheseedconvolution kernels toobtain diversegenerationconvolution kernels.Secondly， thediferentgenerationfunctionsuseddiferenthyperparameterstocontroltheregularizationeffectofthemodelandiproved itsrobustness.Finall，itusedthefeatureapsgeneratedbytheconvoutionalkernls toperformchannelsuflingadconvlutional dimensionalityreductionoperations，whileusinggroupnormalization techniques toimprovethedistributionconsistency offeaturesandenhancetheabilitytocapturecomplex data features.InordertoverifytheefectivenessofCN2Conv，this paper carried outseveral experimentson CIFAR-10，CIFAR-1O0，CIFAR-10-Cand Icons-50datasets.Onthe CIFAR-10-C dataset， the accuracy of ResNet34 using CN2Conv is 8. 22% higher than the standard ResNet34，and 11. 86% higher than MonoCNN. Theresultsshowthatthe accuracyof the CNN modelbasedon CN2Convis beterthanthecomparison method on multiple datasets，and the robustness is significantly improved.

Key words：convolutional neural network （CNN）；cloud-asisted training；less parameter model; robustness

0 引言

根据IDC最新发布的报告《WorldwideGlobalDataSphereIoTDeviceandDataForecast，2023—2027》显示[1，随着物联网技术的持续进步和广泛应用，全球物联网设备的数量预计将从2023年的167亿台激增至2027年的290亿台，其中2024年将达到190亿台。（剩余17424字）