Tampering with only compromised sensors can still fool detectors: PCGAN reaches an 18.69% stealth-impact score on the IEEE 300-bus grid
Synopsis
Working with write access to only a subset of deployed sensor channels and no knowledge of grid topology or detector parameters, PCGAN learns approximate active-power–angle and reactive-power–voltage-magnitude relations from historical measurements and uses adversarial training to generate data-injection perturbations that resemble normal operation. On the IEEE 300-bus system with one year of PJM hourly demand and 10% to 100% compromised sensors, it reaches 84.28% mean evasion of ten machine-learning detectors and 22.31% mean spatial load misperception, for an average stealth-impact score of 18.69% against 5.61% for the strongest competing attack.
Interpretation
Data-injection attacks can now stay hidden and still shift state estimation measurably while the attacker controls only part of the sensors and knows nothing of the grid's internal model. Earlier attack methods traded stealth for impact or impact for stealth, and often assumed the attacker could read and write all bus measurements or knew the detector model. On the IEEE 300-bus system, PCGAN evades ten machine-learning detectors at 84.28% on average, above the next-best BIM at 57.08%; its mean stealth-impact score is 18.69%, versus 5.61% for the strongest competing attack, RL-iGAN.
The attack is built by first learning approximate power-flow sensitivities, then having a generator produce perturbations inside the writable channels that agree with those relations. Those sensitivities are fitted by ridge regression from benign historical measurements of the compromised sensors, requiring no admittance matrix, state estimator or detector parameters. The generator takes historical observations and noise as input, is held to the active-power–angle and reactive-power–magnitude relations by a whitened consistency loss, and a Wasserstein critic pushes perturbations toward benign statistics and timing; evaluation covers 10%, 30%, 60% and 100% compromised sensors.
Perspective
Evaluation uses the IEEE 300-bus system with one year of PJM Interconnection hourly demand and synthetic bus-level load profiles, a placement of 87 PMU/RTU pairs, and compromise levels of 10%, 30%, 60% and 100%. Results cover ten machine-learning detectors plus chi-square and LNR BDD, with physical impact measured as spatial load misperception under nonlinear AC weighted-least-squares state estimation.
The combined score is a seed-matched product of spatial load misperception and detector evasion, not the probability of being stealthy and impactful in the same hour; the potential next-dispatch bias is an estimate, and downstream dispatch was not simulated. The authors state that future work will examine other grids, unseen demand patterns and downstream dispatch consequences.