A Novel Adversarial FDI Attack and Defense Mechanism for Smart Grid Demand-Response Mechanisms

Abstract

This article focuses on enhancing the cybersecurity of cyber-physical systems, with a particular emphasis on the False Data Injection (FDI) attack within the Demand Response (DR) mechanism in smart grids. DR seeks to introduce flexibility in consumers' electricity consumption through dynamic pricing or financial incentives, aiming to optimize the equilibrium between supply and demand. The vulnerability of DR to FDI attacks becomes particularly evident when considering its reliance on accurate demand data. In emphasizing the importance of fortifying DR's security against FDI, the Ensemble and Transfer Adversarial Attack (ETAA) based on Adversarial Machine Learning (AML) techniques is proposed. This method facilitates the injection of false data with reduced detectability by existing neural network-based detection method. With the general framework of ETAA, any gradient-based adversarial attack method can be integrated to achieve enhanced attack transferability across diverse detection models. To counteract such attacks, the training process of detection models is refined through three key steps: Gaussian noise injection, latent feature combination and probability margin enlargement. Evaluation results demonstrate that the ETAA method executes FDI attacks with a higher success rate compared to benchmark methods. Furthermore, defensive training contributes to elevating the performance of detection models, ensuring higher standard accuracy, and reducing the success rate of AML attacks. This paper underscores the critical need to enhance the security of DR mechanisms to mitigate the impact of sophisticated FDI attacks on the robustness of smart grids.