A Cost-Aware Approach to Adversarial Robustness in Neural Networks