On the Adversarial Robustness of Subspace Learning