Recent research has shown that machine learning based Intrusion Detection Systems (IDSs) can be used to detect link layer attacks in 802.11 networks. If these machine learning based detectors are to be considered viable alternatives to conventional detectors, we would need to know more about how robust they are. Do they have any advantages over conventional detectors? Do they have any disadvantages? If so what can be done to mitigate them? Focusing primarily on stealth attacks, cross-platform performance and performance in detecting unknown similar attacks; this work attempts to investigate the questions posed above. The results show that machine learning based detectors, specically Genetic Programming (GP) based detectors, are (or can be) made robust with respect to the above criteria if attention is placed on the presentation of the data used to train them. Role Based MAC (Media Access Control) address mapping is one novel techichnique proposed in this work to achieve this. By showing that machine learning based 802.11 IDSs can be just as robust (if not better) than conventional detectors, this work shows that commercial grade machine learning based 802.11 IDSs can be built.