Point Adjustment Makes Random Scores Look Competitive on Power-Plant Control Data: An Evaluation Study on HAI 21.03

On HAI 21.03, a power-generation industrial control testbed with 50 labelled attacks, uniformly random scores reach F1 = 0.71 under point adjustment, close to an Isolation Forest (0.77), while their point-wise F1 is 0.04. Requiring that 10% of an attack be flagged (PA%K, K = 10) drops the random control to 0.25. With thresholds set from held-out normal data, random scores produce 84.6 false alarms per day while a PCA detector finds 45 of 50 attacks with a 21 s median delay and 2.4 false alarms per day. Code, results, a 4-page technical report and an A1 poster.