The content discusses the problem of learning minimal neural activation pattern (NAP) specifications for neural network verification. Key points:
Specifications play a crucial role in neural network verification, as they define the precise input regions to be verified. Recent research suggests using NAPs as specifications, but focuses on computing the most refined NAPs, which are often limited to small regions in the input space.
The authors study the problem of finding the minimal (coarsest) NAP specification that is sufficient for formal verification of the network's robustness. This is important as minimal NAP specifications can expand verifiable bounds and provide insights into which neurons contribute to the model's robustness.
The authors propose exact approaches (Refine and Coarsen) that leverage verification tools to find minimal NAP specifications, as well as approximate approaches (Sample_Refine, Sample_Coarsen, Adversarial_Prune, Gradient_Search) that efficiently estimate minimal NAPs without making calls to the verification tool.
The authors also introduce a method to estimate the volume of regions corresponding to NAP specifications, which helps understand the volumetric change between minimal and refined NAP specifications.
Experiments show that minimal NAP specifications involve much smaller fractions of neurons compared to the most refined NAPs, and they significantly expand the verifiable bound by several orders of magnitude.
toiselle kielelle
lähdeaineistosta
arxiv.org
Tärkeimmät oivallukset
by Chuqin Geng,... klo arxiv.org 04-09-2024
https://arxiv.org/pdf/2404.04662.pdfSyvällisempiä Kysymyksiä