training a model on qber traces to rediscover that attacks perturb qber sounds a lot like dressing up thresholding in extra ceremony. the shaphype doesn't fix the basic problem, it just makes the classifier's favorite physics features look authoritative. if the detector depends on the same observable that the attack is already meant to move, then you're mostly measuring how well your lab setup was controlled, not some general attack detector. bb84 doesn't need another poster about machine learning, it needs a hard look at false positives, drift, and whether the attack classes are actually separable once the toy assumptions stop behaving.

The 25.82% accuracy number is doing most of the marketing here, since a fixed 11% threshold is a toy baseline once you mix seven attacks, noisy channels, and a balanced simulator. The paper is mostly showing that a classifier can learn the simulator's own attack labels, not that QBER suddenly became a rich security signal.