On the occasion of the 111th birthday of Edith Piaf’s birth, we are happy to release CAISAR version 5.0.
The release source is available at our public forge and, as always soon on opam.
This release provided the following enhancements
Local robustness consist on asserting whether a neural network output changes given a perturbtation on the input. While important, this property fails to describe behaviour on the whole input domain, which ampers the expressiveness of specifications.
CAISAR 5.0 has support for global robustness properties, through the lens of Confidence-Based Properties. It is now possible to specify and verify global robustness properties with CAISAR. This opens the possibility to specify properties on a whole input domain.
Thanks to Guilhem Ardouin for this work. This feature is described in a publication (in French): La Confiance avec le Contrôle: Spécification d’hyperpropriétés sur réseaux de neurones.
Added the prover-precision option. This option allows to automatically adjust
the tradeoff between precision and speed for all of CAISAR supported provers.
This option opens the way for further automated configuration for CAISAR, as to
ease the burden of verifying a property.
We also fixed the support for $\alpha-\beta$-CROWN, updated prover versions, bumped dependencies and fixed more bugs.