Dynamically Computing Adversarial Perturbations for Recurrent Neural Networks