This is a real distance-based outlier detector, not a cartoon: it estimates
a robust centroid (coordinate-wise median, immune to being dragged by the
outliers it's trying to find) and a robust spread (median absolute
deviation, MAD) from the current point cloud, then scores every point.
centroid = median(points)
spread = MAD(points, centroid)
score(x) = |x - centroid| / spread
flag if score > sensitivity
- Data points — total sample size drawn each reset.
- Cluster tightness — standard deviation of the normal-behaviour
Gaussian; tighter clusters sharpen the boundary.
- Anomaly injection rate — fraction of points deliberately drawn
from a uniform outlier distribution scattered around the canvas.
- Detection sensitivity — the MAD-multiple threshold; lower catches
more (and more false positives), higher misses subtle outliers.
Because we know which points were actually injected as outliers, the
Accuracy stat compares the algorithm's verdict against ground truth —
the same evaluation used for real fraud-detection and intrusion-detection
pipelines scoring deviation from a learned baseline.