Since its introduction in 2016, federated learning (FL) has been used in multiple domains, such as intrusion detection. However, FL literature shows that the heterogeneity of most real-world FL applications makes it difficult for clients to converge in a suitable global model. Furthermore, as a collaborative system, FL is vulnerable to attacks, such as model poisoning. While strategies have been identified in the literature, they often rely on the assumption that the data distribution among participants is homogeneous. In this paper, we review the current challenges in clustering and adversarial mitigation in heterogeneous FL, and propose different strategies to address them. Namely, we present a cross-evaluation framework for exhaustive gathering, and a set of algorithmic countermeasures based on principal component analysis. We show preliminary results of our clustering mechanism, which validates the effectiveness of the cross-evaluation framework.