Exploiting Verified Neural Networks via Floating Point Numerical Error

03/06/2020

∙

by   Kai Jia, et al.

∙

0

∙

share

We show how to construct adversarial examples for neural networks with exactly verified robustness against ℓ_∞-bounded input perturbations by exploiting floating point error. We argue that any exact verification of real-valued neural networks must accurately model the implementation details of any floating point arithmetic used during inference or verification.