基于人工智能技术的配电系统图像分类模型安全评估体系构建与测试研究  

Research on the construction and testing of security assessment system of image classification modelfor power distribution system based on artificial intelligence technology

作　　者：郑州 钱健 李扬笛 谢炜 马腾 周晨曦 赵志超 ZHEN Zhou;QIAN Jian;LI Yangdi;XIE Wei;MA Teng;ZHOU Chenxi;ZHAO Zhichao(State Grid Fujian Electric Power Co.,Ltd.,Electric Power Science Research Institute,Fuzhou,Fujian,350000,China)

机构地区：[1]国网福建省电力有限公司电力科学研究院,福州350000

出　　处：《自动化与仪器仪表》2024年第9期339-344,共6页Automation & Instrumentation

摘　　要：深度学习是利用大型数据集的高效训练算法,在新型智能电力系统中得到了广泛应用。然而,目前电力系统人工智能模型的深度神经网络训练阶段不完善,容易受到对抗性样本的影响,导致深度神经网络出现错误分类。因此,在新型电力系统中,确保基于配电系统图像多分类深度神经网络模型的安全性至关重要。本文首先确定了针对配电系统图像多分类的深度神经网络的对抗空间,并根据对模型的输入和输出之间映射的精确理解来制作对抗者样本。然后,定义了一个对抗攻击能力指标来评估不同样本类别对对抗性扰动的安全性。最后,通过定义良性输入和目标分类之间的距离的预防性措施,实现了对对抗性样本的防御。Deep learning is an efficient training algorithm using large data sets,which is widely used in new intelligent power systems.However,the current deep neural network training stage of artificial intelligence models for power systems is imperfect and vulnerable to adversarial samples,which leads to misclassification of deep neural networks.Therefore,it is crucial to ensure the security of multi-classification deep neural network models based on distribution system images in new power systems.In this paper,we firstly determine the adversarial space of a deep neural network for multi-classification of distribution system images and produce adversary samples based on an accurate understanding of the mapping between the inputs and outputs of the model.Secondly,an adversarial attack capability metric is defined to evaluate the security of different sample classes against adversarial perturbations.Finally,defense against adversarial samples is achieved by defining a precautionary measure of the distance between benign inputs and target classifications.

关 键 词：智能电力系统 对抗性样本 深度神经网络 配电系统图像多分类 安全性 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]