Robust Optimal Classification Trees against Adversarial Examples

Decision trees are a popular choice of explainable model, but just like neural networks, they suffer from adversarial examples. Existing algorithms for fitting decision robust against examples greedy heuristics and lack approximation guarantees. In this paper we propose ROCT, collection methods to train that optimally user-specified attack models. We show the min-max optimization problem arises in learning can be solved using single minimization formulation with 0-1 loss. such formulations Mixed-Integer Linear Programming Maximum Satisfiability, which widely available solvers optimize. also present method determines upper bound on accuracy any model bipartite matching. Our experimental results demonstrate existing achieve close optimal scores while ROCT achieves state-of-the-art scores.