DeepLR: Deep Learning Assisted Load Redistribution Attack on Cyber-Physical Power System

In this work, we propose DeepLR, a data-driven Load Redistribution (LR) attack model on cyber-physical power systems. Prior attack models in this domain are fundamentally inadequate, as they fail to address or offer only partial/isolated solutions to the three critical challenges of effective attack modeling: (i) high computation time, (ii) reliance on confidential (complete/partial) topology information, and (iii) the need to manipulate all loads in the grid. Significantly diverging from prior attack models, this paper presents DeepLR, a Deep Learning (DL) assisted attack model designed to address these three critical challenges of effective attack modeling in a single unified framework. We first reformulate the conventional LR attack problem into a DL formulation, thereby eliminating reliance on confidential (complete/partial) topology information in real-time. Second, we propose a data-driven approach to identify a potential attacking region (characterized by fewer lines and loads), which enables attackers to focus exclusively on manipulating a subset of loads within this region rather than the entire grid. Through numerical experiments and comparative analysis with prior attack models, we demonstrate two critical findings: (i) an equally potent LR attack can be constructed without access to confidential topology information in real-time, and (ii) targeting the identified attacking region, even with fewer manipulations, results in a more severe impact on grid operations. The case study on the IEEE 14 bus system and the IEEE 39 bus system demonstrates the efficacy of the proposed DeepLR approach.