Anomaly detection IDS, sometimes, are designed to prevent mimicry attacks.
After the algorithm has done the clustering, there might be few and small clusters. The attacker will have problems with generating malicious samples that can be clustered as normal because it is more difficult for her/him to find the space where the "normal" clusters are.
What can be done to make a mimicry attack successful? How to put the malicious sample inside the normal clusters?