On the Adversarial Inversion of Deep Biometric Representations