SLA$^2$P: Self-supervised Anomaly Detection with Adversarial Perturbation