基于BiLSTM与自注意力机制生成对抗网络的GSA防护方法

Defense Method for Smart Grid GPS Spoofing Attack Based on BiLSTM and Self-attention Mechanism Generative Adversarial Network

相量测量单元(phasor measurement unit,PMU)是智能电网的重要组成部分,能精准同步采集电力数据。由于PMU使用全球定位系统(global positioning system,GPS)提供时间同步参考,容易遭受GPS欺骗攻击(GPS spoofing attack,GSA),影响正常的数据采集。现有GSA防御方法的修复精度较低且需要额外的硬件成本。为了解决上述问题,提出一种基于双向长短期记忆网络与自注意力机制生成对抗网络的GSA防护方法。首先,提出一种改进的带梯度惩罚的Wasserstein生成对抗网络(Wasserstein generative adversarial network with gradient penalty,WGAN-GP)模型,重新设计原有生成器和判别器的网络架构,并在生成器和判别器中分别引入双向长短期记忆网络以及自注意力机制,提升模型的生成性能和鉴别能力。其次,基于所提出的WGAN-GP模型,构建了一种GSA防御模型,其包含攻击检测网络和数据修复网络2个模块,分别用于检测智能电网GSA和修复受损的PMU测量数据。最后,在IEEE-39总线系统中模拟GSA攻击,并在相应的数据集验证方法的有效性。结果表明,与现有方法对比,所提方法在大部分性能指标上取得了领先的性能。

Phasor measurement unit(PMU)plays a crucial role in smart grids,enabling precise synchronized acquisition of electric power data.Due to the use of the global positioning system(GPS)for time synchronization,the PMU is vulnerable to GPS spoofing attack(GSA),which impacts the normal data acquisition.The existing GSA defense methods have low restoration accuracy and require additional hardware costs.To address the aforementioned issues,this paper proposes a GSA defense method based on bidirectional long short-term memory(BiLSTM)network and self-attention mechanism generative adversarial network.Firstly,an improved WGAN-GP model is proposed to redesign the network architecture of the generator and discriminator,and the BiLSTM network and self-attention mechanism are incorporated into the generator and discriminator to enhance the model's generative performance and discriminative ability.Secondly,based on the proposed WGAN-GP model,a GSA defense model is constructed,which includes two crucial modules:an attack detection network and a data restoration network that are employed to detect the smart grid GSA and repair the compromised PMU measurement data,respectively.Finally,We simulated GSA attacks in the IEEE-39 bus system and validated the effectiveness of the proposed method on the corresponding dataset.The results show that compared to existing methods,the proposed approach outperforms in most performance indicators.