This library contains implementations of various methods to soundly verify deep neural networks. In general, we verify whether a neural network satisfies certain input-output constraints. The verification methods are divided into five categories:
Reference: C. Liu, T. Arnon, C. Lazarus, C. Barrett, and M. Kochenderfer, "Algorithms for Verifying Neural Networks," arXiv:1903.06758.
To download this library, clone it from the julia package manager like so:
(v1.0) pkg> add https://github.com/sisl/NeuralVerification.jl
Please note that the implementations of the algorithms are pedagogical in nature, and so may not perform optimally. Derivation and discussion of these algorithms is presented in the survey paper linked above.
Note: At present,
Duality do not work in higher dimensions (e.g. image classification).
This is being addressed in #9
The implementations run in Julia 1.0.
using NeuralVerification solver = BaB()
nnet = read_nnet("examples/networks/small_nnet.nnet") input_set = Hyperrectangle(low = [-1.0], high = [1.0]) output_set = Hyperrectangle(low = [-1.0], high = [70.0]) problem = Problem(nnet, input_set, output_set)
julia> result = solve(solver, problem) CounterExampleResult(:violated, [1.0]) julia> result.status :violated
For a full list of
Solvers and their properties, requirements, and
Result types, please refer to the documentation.
12 days ago