脉冲神经网络对抗样本攻击与防御综述OA

With the wide deployment and application of spiking neural networks,their security problems have become more and more obvious,especially the threat from adversarial example attacks.Therefore,the adversarial example attack methods and defense measures in spiking neural networks are investigated.Firstly,the adversarial sample attack methods are studied,and the software layer is sorted out from the gradient attack,migration learning attack,encoding perturbation attack and sensor attack.The hardware layer deals with power injection attack,side channel attack and Trojan horse attack.Secondly,the defense measures of adversarial samples are studied.The defense measures of the software layer are sorted from adversarial training,input filtering,improved coding,feature network analysis and model fusion.The defense mea-sures of hardware layer are discussed from two parts of circuit optimization and security framework.Then,the application of adversarial examples in model security research and CAPTCHA anti-identification is discussed.Finally,the current challenges and future prospects are put forward and the whole paper is summarized.