Adversarial training of gradient-boosted decision trees