PV2 ⊧ GDVB: Generation of DNN Verification problem Benchmarksgiven a problem and constraints, generates benchmarks

Application domain/field

• Machine learning
• Deep neural networks
• Benchmarking

Type of tool

Benchmark/test generator

Expected input

• Seed problem $<n_s,{\phi }_s>$
• Set of factors $F$
• Constraints $\Gamma$
• Coverage goal $t$

Format:

A .toml configuration file

Expected output

A benchmark of DNN verification problems

Internals

GDVB generates benchmarks for the verification of neural networks. The benchmarks that it generates seeks to cover variations that influence verifier performance. Uses R4V and DNNV.