Answers are grounded in this paper and cite their sources.
False-data-injection (FDI) detectors for power grids are usually demonstrated against attacks large enough to trip a state estimator. The attacks that matter operationally are the opposite kind: small, slow, and sized to stay inside the noise. We ask a simpler and, we argue, more useful question than "does the detector work": how large must an injection be before each detector can see it at all? Using the UCI Electrical Grid Stability Simulated Data (10,000 operating points of a four-node decentralised-control grid), we corrupt the reported features of a single grid participant with scale, shift, and slow-ramp (drift) injections at magnitudes from 1% to 50%, and we measure detection rate against attack magnitude for four detectors fitted on clean data only: a supervised stability classifier (symptom: prediction flips), an Isolation Forest, a physics-residual detector that predicts the participant's reports from the other nodes and scores the Mahalanobis residual, and an EWMA process-control chart on that residual stream. Because the simulated data obeys an exact power-balance identity, the residual detector is trivially perfect on the as-recorded data; we therefore report a second, headline condition in which every feature carries sensor noise of 1% of its standard deviation and all thresholds are recalibrated at the same nominal 5% false-positive rate. Under sensor noise, the detection boundaries separate sharply. The physics-residual detector reaches 88% detection for a 1% scale injection and 94% for a 5% shift injection; the EWMA chart accumulates evidence and reaches a 96% alarm rate for a 2% shift injection that the per-row residual detector catches only 30% of the time, and it alarms 14 samples into a drift attack whose instantaneous magnitude is still below 0.4%. The Isolation Forest never exceeds 38% detection even at a 50% injection, and the stability classifier flips at most 11% of its predictions at 50% — a 50% corruption of one participant's reports costs it only about 6.5 accuracy points. Generic, physics-blind monitoring of this stream is, in a quantifiable sense, silent; a one-line physical invariant is not, but only down to the noise floor that carries it.
No takes yet. Share an insight, caveat, or question.
Saagar Sengupta (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: