Diffusion-based Random Attack Generators for Interdiction Analysis in Electric Power Systems

A diffusion-based learning approach is proposed to train a Diffusion-based Random Attack Generator (DRAG) in the context of interdiction analysis for electric power systems. The DRAG is trained on solutions to an interdiction optimization problem that identifies worst-case attacks (or interdictions) that disrupt the adversary's operation. It is conditioned on the continuous demand of the power system and intends to generate random severe sparse attacks in a discrete space representing network components to be attacked (or interdicted). The proposed approach introduces a regularization term in the loss function that represents the maximum-load delivery problem, effectively rendering the loss minimization problem a variant of the interdiction problem itself. The regularized loss improves the performance of the DRAG and allows it to produce severe attack samples not seen in training. The resulting DRAG is able to generate severe attacks more quickly than solving the interdiction optimization problem. It additionally mimics the random behavior of a true adversary, making it a useful tool for analyzing power system defenses. Examples are provided on the IEEE 300 bus test case.