DeepZ is a method for local robustness verification, based on zonotopes, a type of convex relaxations, and abstract transformations. While affine transformations and convolutions can be represented exactly using this framework, a sound over-approximation has to be used to approximate the ReLU function. The abstract transformer used in DeepZ is parameterized by one parameter per hidden unit. In this work we propose an optimization strategy for this parameterization to increase the maximum verifiable image perturbation for both fully-connected and convolutional network architectures.
DeepZ is a method for local robustness verification, based on zonotopes, a type of convex relaxations, and abstract transformations. While affine transformations and convolutions can be represented exactly using this framework, a sound over-approximation has to be used to approximate the ReLU function. The abstract transformer used in DeepZ is parameterized by one parameter per hidden unit. In this work we propose an optimization strategy for this parameterization to increase the maximum verifiable image perturbation for both fully-connected and convolutional network architectures.