Improving the robustness of decision trees in security-sensitive setting