[2405.17130] Exploiting the Layered Intrinsic Dimensionality of Deep Models for Practical Adversarial Training