基于联帮学习的深度学习模型保护算法研究.pdfVIP

浙江大学硕士学位论文摘要

3.在几种常见的联邦学习应用的任务中进行了对照实验。若没有作恶正常

训练的情况下，添加冗余结构的模型性能没有太大改变。若盗用方节点直

接盗用复杂模型进行重新训练，模型预测精度会明显降低，训练时间增加，

在复杂任务中模型甚至难以收敛，从而增加了计算节点盗用模型结构的

成本，降低了模型被盗用的风险。

关键词：深度学习，联邦学习，神经网络，隐私保护

ii

浙江大学硕士学位论文Abstract

Abstract

Inthecurrentresearchboomoffederallearning,mostoftheexistingresearch

worksfocusonhowtoprotectprivatedatafromthirdpartiesinmachinelearning

training,orexplorehowtosharedatafeaturesindifferenttaskscenarios,etc.,butlittle

attentionispaidtomodelprotection.

Whileresearchhasrecentlyemergedtoaddresstheperformancesecurityof

federatedlearningmodels,suchasmodelupdatepoisoninganddatapoisoningattacks,

theparasiticevilscenariohasnotbeenspecificallystudiedandeffectivelyaddressed.

Parasiticevilreferstotheparasiticbehaviorinwhichthecomputingnodedirectlysteals

thetrainingmodeldistributedbytheaggregationnodeanddoesnotcausedamagetothe

overallfederationlearningarchitecture.Toaddresstheproblemofprivacyleakage

causedbythisparasiticbehavior,thispaperpresentsthefirstsolutiontoprotectthe

modelsprovidedbyaggregationnodestoreducetheriskoftheftbycomputationalnodes.

Inthispaper,theresearchiscarriedoutfromtwoperspectivesofmodelweight

protectionandmodelstructureprotection,respectively,andthemainresearchcontents

andcontributionsareasfollows：

1.Forthescenarioinw