【金陵读书】人工智能：AI时代的网络安全（2022.6.18）( 三 ) 【金陵读书】初探AI安全（20

文章图片
总结
人工智能作为第四次工业革命的核心应用技术之一，正以高度自动化和自主性的特性，创造出更为巨大的利益。与此同时，伴随着“AI-as-a-Service”的普及， AI应用所需专业技能与知识门槛的降低，一定程度上增加了AI被恶意使用和滥用的可能性，进一步扩大了以人工智能为载体的安全威胁攻击面。简言之，人工智能在塑造一个更广阔前景的同时，也因攻击者的日趋深入，而成为诱发新数字、社会和政治威胁的额外动力。
归根结底，对安全从业人员和安全公司而言，人工智能对信息安全来说无疑是一把双刃剑：其带来巨大技术革新的同时，亦带来了前所未有的全新挑战。一方面，由于人工智能是一门包含了计算机科学、数学、生物等方面的综合性学科，如何掌握并使用人工智能，并将其应用在安全相关领域应成为传统信息安全从业人员以及公司的关注重点。同时，针对数据的采集以及数据隐私问题更应该被严格对待，将网络安全和数据保护放在首位，这要求相关部门、企业、研究人员进行协作，在保证数据的安全及隐私的同时，促进相关创新、法规建设、提高认识和开展研发活动，共同创建人工智能安全发展的健康环境。
引用文献：
[1]MasteringthegameofGowithdeepneuralnetworksandtreesearch.
[3]Krizhevsky,Alex,IlyaSutskever,andGeoffreyE.Hinton."Imagenetclassificationwithdeepconvolutionalneuralnetworks."Advancesinneuralinformationprocessingsystems25(2012):1097-1105.
[4]He,Kaiming,etal."Deepresiduallearningforimagerecognition."ProceedingsoftheIEEEconferenceoncomputervisionandpatternrecognition.2016.
[5]Vaswani,Ashish,etal."Attentionisallyouneed."arXivpreprintarXiv:1706.03762(2017).
[6]Mnih,Volodymyr,etal."Playingatariwithdeepreinforcementlearning."arXivpreprintarXiv:1312.5602(2013).
[7]Brown,Noam,etal."Combiningdeepreinforcementlearningandsearchforimperfect-informationgames."arXivpreprintarXiv:2007.13544(2020).
[8]Brown,TomB.,etal."Languagemodelsarefew-shotlearners."arXivpreprintarXiv:2005.14165(2020).
[9]CLIP:ConnectingTextandImages
[10]LearningtoConfuse:GeneratingTrainingTimeAdversarialDatawithAuto-Encoder.JiFeng,QizhiCai.ZhiHuaZhou.
[11]McMahan,Brendan,etal."Communication-efficientlearningofdeepnetworksfromdecentralizeddata."ArtificialIntelligenceandStatistics.PMLR,2017.
[12]Yang,Qiang,etal."Federatedmachinelearning:Conceptandapplications."ACMTransactionsonIntelligentSystemsandTechnology(TIST)10.2(2019):1-19.
[13]Goodfellow,IanJ.,JonathonShlens,andChristianSzegedy."Explainingandharnessingadversarialexamples."arXivpreprintarXiv:1412.6572(2014).
[14]Madry,Aleksander,etal."Towardsdeeplearningmodelsresistanttoadversarialattacks."arXivpreprintarXiv:1706.06083(2017).
[15]Carlini,Nicholas,andDavidWagner."Towardsevaluatingtherobustnessofneuralnetworks."2017ieeesymposiumonsecurityandprivacy(sp).IEEE,2017.
[16]Moosavi-Dezfooli,Seyed-Mohsen,AlhusseinFawzi,andPascalFrossard."Deepfool:asimpleandaccuratemethodtofooldeepneuralnetworks."ProceedingsoftheIEEEconferenceoncomputervisionandpatternrecognition.2016.
[17]Ma,Chen,LiChen,andJunhaiYong."MetaSimulator:SimulatingUnknownTargetModelsforQuery-EfficientBlack-boxAttacks."arXivpreprintarXiv:2009.00960(2020).
[18]Yin,Hongxu,etal."SeethroughGradients:ImageBatchRecoveryviaGradInversion."arXivpreprintarXiv:2104.07586(2021).
[19]GenerativeAdversarialNetworks.IanJ.Goodfellow,JeanPouget-Abadie?,MehdiMirza,BingXu,DavidWarde.SherjilOzair,AaronCourville,YoshuaBengio.
[20]Weaponizingdatascienceforsocialengineering:AutomatedE2EspearphishingonTwitter.
[21]DeepLearningforCyberVulnerabilityDiscovery:NGTFProjectScopingStudy.deVelO,HubczenkoD,KimJ,MontagueP,XiangY,PhungD,ZhangJ,MurrayT,LeT,WenS,LiuS,NguyenV,LinG,NguyenK,LeT,NguyenT,NockR,QuL.