Kernel Learning with Adversarial Features: Numerical Efficiency and Adaptive Regularization