基于改进EfficientNet的电力资产信息数据流量异常检测的应用  

An Anomaly Detection Scheme for Power Asset Information Data Flow Based on Improved EfficientNet

作　　者：李泽科 郭久煜 邓春荣 张章学 LI Zeke;GUO Jiuyu;DENG Chunrong;ZHANG Zhangxue(State Grid Fujian Electric Power Co.,Ltd.,Fuzhou Fujian 350003,China;Longyan Power Supply Co.,Ltd.,State Grid Fujian Electric Power Co.,Ltd.,Longyan Fujian 364100,China;Fujian Strait Information Technology Co.,Ltd.,Fuzhou Fujian 350003,China)

机构地区：[1]国网福建省电力有限公司,福建福州350003 [2]国网福建省电力有限公司龙岩供电有限公司,福建龙岩364100 [3]福建省海峡信息技术有限公司,福建福州350003

出　　处：《信息安全与通信保密》2024年第5期42-53,共12页Information Security and Communications Privacy

基　　金：中央引导地方科技发展专项(2021L3032)。

摘　　要：针对传统异常流量检测方法在面对复杂多样的新型网络攻击时存在的数据特征提取困难、准确率低、误报率高和运行成本高等问题,提出了一种基于EfficientNet与贝叶斯超参数优化的电力资产信息流量异常检测方法,通过将流量数据序列转换为二维图像,有效提取网络流量异常特征并进行分类,实现高准确度和效率。在KDD-NSL、CIC-IDS2017数据集上,相较于VGG19、Xception、ResNet50,该方法分别提升最高6.7%的准确率和8.1%的F1值,表明了该方法的显著优势和实用价值。Conventional traffic anomaly detection methods have some problems when facing complex and diverse new network attacks,such as difficulties in data feature extraction,low accuracy,high false positives,and high operation cost.To address these problems,this paper proposes an anomaly detection method for power asset information data flow based on EfficientNet and Bayesian hyperparameter optimization.By transforming traffic data sequences into 2D images,the method could effectively extract features of network traffic anomalies and carry out classification to achieve high accuracy and efficiency.On KDD-NSL and CIC-IDS2017 datasets,compared with VGG19,Xception,and ResNet,the proposed method demonstrates improvements of up to 6.7%in accuracy and 8.1%in F1 score,highlighting its significant advantages and practical value.

关 键 词：贝叶斯超参数优化 网络攻击 异常流量检测 智能变电站 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]