Expand description
Detection metrics for the comparative benchmark.
Pure functions over labels and scores: a binary confusion matrix (positive = stego), the rank-based ROC AUC (Mann-Whitney, tie-aware), and the ROC curve points. These are deliberately storage- and tool-agnostic so the same code scores Stegcore’s ensemble, any single detector, or an external comparator.
Structs§
- AucInterval
- An AUC with the uncertainty that belongs beside it.
- Confusion
- Binary classification counts. Positive is “stego”.
- Operating
Point - The point on the curve that
tpr_at_fpractually reported. - RocPoint
- One vertex of the ROC curve, including the score that produced it.
Enums§
- Finding
- Something about a set of scores that makes the number beside it misleading.
- Tally
Error - Why a confusion count could not be taken honestly.
Constants§
- Z_95
- 1.959964, the two-sided normal quantile for 95 per cent.
Functions§
- findings
- Every
Findingthat applies to these scores and labels. - fpr_
resolution - The smallest false-alarm rate a corpus of this many clean images can show.
- operating_
point tpr_at_fpr, plus the rate the answer actually came from.- roc_auc
- Rank-based ROC AUC (equivalent to the Mann-Whitney U statistic), tie-aware via average ranks.
- roc_
auc_ interval - AUC with a confidence interval, by DeLong’s method.
- roc_
curve - ROC curve as
(fpr, tpr)points, swept from the highest score downward (predicted positive whenscore >= threshold). Begins at(0, 0)and ends at(1, 1). - roc_
points - The ROC curve as vertices carrying the threshold each one sits at.
- tpr_
at_ fpr - The best true-positive rate this detector reaches without exceeding
max_fpr.