Distribution-based Artificial Anomaly Generation

Wei Fan, et. al propose an interesting technique for generation of artificial anomaly data. The technique is independent of the learner.
The possible issue is that they generate anomalies "dimension by dimension" - they treat each dimension independently. They propose that anomalies could consider multiple dimensions concurrently -or have different weights on dimensions.
They use this technique to "define decision boundaries that separate the given class labels".

I plan to use this technique to train my Anomaly detector with the DARPA dataset.