Verifying Properties of Gradient Boosting Model
17 July 2019
Verifying properties of machine learning models is a critical challenge. Gradient Boosting is an important technique which allows weak learners to be combined into strong ones. This work is about verifying properties for Gradient Boosting models. Our main contribution is the encoding of such models as SMT expression which enables the verification of properties on such models. Specifically, we target the critical property of robustness to adversarial perturbations. Our evaluation show that for some datasets the models are much more robust than for others. We also show that it is easier to verify robustness for models with smaller depth rather than deep ones which are known to suffer from overfitting. To the best of our knowledge, our work is the first to verify properties of Gradient Boosting models.
