基于Transformer编码器的智能电网虚假数据注入攻击检测  

FALSE DATA INJECTION ATTACKS DETECTING BASED ON TRANSFORMER ENCODER IN SMART GRID

作　　者：陈冰 唐永旺 Chen Bing;Tang Yongwang(Department of Electrical and Electronic Engineering,Henan University of Technology Luohe Institute of Technology,Luohe 462000,Henan,China;School of Information System Engineering,PLA Support Force Information Engineering University,Zhengzhou 450002,Henan,China)

机构地区：[1]河南工业大学漯河工学院电气电子工程系,河南漯河462000 [2]中国人民解放军战略支援部队信息工程大学信息系统工程学院,河南郑州450002

出　　处：《计算机应用与软件》2022年第7期336-342,共7页Computer Applications and Software

摘　　要：针对当前基于循环神经网络的智能电网虚假数据注入攻击(False Data Injection Attacks, FDIA)检测方法无法同时利用量测样本中前后参数信息和样本间参数关联关系的问题,提出一种基于Transformer编码器的FDIA检测框架。对连续时间样本数据进行归一化处理,结合相对位置信息得到连续时间样本向量。引入Transformer编码器,通过多头自注意力机制计算长距离依赖关系,得到连续时间样本的特征表示。将该特征表示输入到全连接神经网络层和Softmax层,输出后一时刻样本受到注入攻击的概率。在IEEE 14-bus和IEEE 30-bus中的仿真实验结果表明该方法切实可行,相较于次优结果,准确率平均提高7.41%,正报率平均提高4.51%,误报率平均降低60.99%。The current detection method based on the recurrent neural network for false data injection attacks(FDIA)in smart grid cannot simultaneously use the front-back parameter information in single measurement sample and the parameter relationship between the samples.To solve this problem,we propose the FDIA detection method based on transformer encoder.The continuous time sample data was normalized,and the continuous time sample vector was obtained by combining the relative position information.We introduced the transformer encoder to calculate the long-distance dependence through the multi-head self-attention mechanism,and then gained the feature representation of continuous time sample.The feature representation was input into the neural network layer and Softmax layer,and the probability of injection attack at the next moment was output.The simulation results in IEEE 14-bus and IEEE 30-bus show that this method is feasible.And compared with the sub-optimal results,the average accuracy rate is increased by 7.41%,the positive report rate is increased by 4.51%,and the false alarm rate is decreased by 60.99%.

关 键 词：Transformer编码器 连续时间 多头注意力 智能电网 虚假数据 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]