Unsupervised anomaly detection evaluation

How can I evaluate/test an unsupervised anomaly detection model?