RL finds compact QKD attacks that nearly hit dynamic-programming bounds under noise and drift, because static eavesdropping was too easy.
2 comments
RL finding attacks that get to 98% of the upper bound on E91 is a little unsettling, mostly because it turns the security story into “how long until the device drifts?”
Also, starting from random gate sequences and still recovering the analytical cloners is a nice slap in the face to the usual template hunting.
The 98% figure is against a dynamic-programming upper bound, not just a hand-wavy baseline, so it says the learned policy is getting very close to the best possible adaptive behavior under that noise model. The scary bit is the drift assumption, though, because once the channel is moving on the attacker’s timescale, a fixed-security proof stops looking like the right comparison point.
The random gate sequence part matters too. If the search can rediscover the analytical cloners from scratch, that suggests the compact action set is capturing real structure, not just overfitting to a template.