False data injection attack sample generation using an adversarial attention-diffusion model in smart grids

Abstract

<p>A false data injection attack (FDIA) indicates that attackers mislead system decisions by inputting false or tampered data into the system, which seriously threatens the security of power cyber-physical systems. Considering the scarcity of FDIA attack samples, the traditional FDIA detection models based on neural networks are always limited in their detection capabilities due to imbalanced training samples. To address this problem, this paper proposes an efficient FDIA attack sample generation method by an adversarial attention-diffusion model. The proposed scheme consists of a diffusion model and a GAN model with an attention mechanism (ATTGAN). First, the forward diffusion of the diffusion model was used to add noise to the real data while injecting the attack vector. Then, the ATTGAN model was trained to effectively focus on the information of power grid measurements and topological nodes, while weakening irrelevant information. In the reverse diffusion process, the trained ATTGAN model was combined to predict the noise, and it was further iterated forward step by step and denoised in this process. Finally, a large number of efficient FDIA attack samples can were generated. Extensive experiments have been carried out on IEEE 14, IEEE 39, and IEEE 118 bus systems. The experimental results indicate that the generated attack samples outperform existing state-of-the-art schemes in terms of evasion detection capability, robustness, and attack strength.</p>