Mixed rounds, and a detector that was reading its own reference¶
Extends the entry above to the deployment question: not "was this named client in the round" with the other participants drawn from other organizations, but "did this organization take part" with the round drawn from every client. The organization's clients are split once into the attacker's reference and the participants it may contribute, so no round is scored against itself.
Averaged over eight random reference splits, each drawn from a stream independent of the rounds, with the standard deviation across splits:
| target | round of 4 | round of 8 | round of 16 |
|---|---|---|---|
| OpenStack, 1 of its clients present | AUC 0.659 (sd 0.070) | 0.688 (0.062) | 0.753 (0.086) |
| OpenStack, 2 of its clients present | 0.771 (0.138) | 0.828 (0.090) | 0.887 (0.084) |
| Qt, 1 of its clients present | 0.668 (0.061) | 0.721 (0.051) | not measurable |
| Qt, 2 of its clients present | 0.765 (0.063) | 0.852 (0.075) | not measurable |
Detection rises with how many of the organization's clients are in the round, and rises with the round rather than falling: a larger round dilutes the target's share and averages the other participants' noise away at the same rate. Qt's rounds of sixteen are not measurable here, since with 22 of the 34 clients it leaves too few outsiders to fill one without it.
The ordering is reliable and the difference is small. Behind OpenStack's 0.887 the two classes' mean scores are 0.7151 and 0.7065: one part in a hundred. An aggregate does not shout its participants; it leans, consistently, and the next entry is what that leaning is worth to an attacker who watches more than one round.
(The first run of this table read 0.765, 0.726 and 0.725 on OpenStack's one-client row, from a single reference split taken in listing order; see the correction below.)
First run of this said AUC 1.000, and that was a defect in the detector. The two classes were scored against references of different sizes, eleven clients for rounds holding the target and six for rounds without, and a larger reference has a more central direction, so the comparison read reference size rather than membership. One fixed reference for both classes gives the numbers above. A test now pins the split.