Modern machine learning applications deal with various types of data, which might originate from different entities and/or spread across different geographical locations. In particular, anomaly detection, relies on combining diverse data from different sources to generalize well. However, transferring data to a centralized location for further processing is not always possible due to data protection, usage, and ownership restrictions. This results in data silos that cannot be merged and represent a serious impediment for most anomaly detection tasks dealing with sensitive data. Federated learning provides a privacy-preserving solution that allows training machine learning models across multiple distributed clients holding local data without exchanging them. While preserving privacy, such solution is usually associated with loss in the predictive performance compared to centralized training. In this thesis, we provide a comprehensive evaluation of federated learning when applied to anomaly detection for different label availability scenarios. We investigate the effect of federated learning on the predictive performance for various applications. For this purpose, federated models are compared to models trained using only the locally available data, to models trained on centrally aggregated data, and to centralized models trained on aggregated synthetic data generated by each client individually. We also investigate the effect of amount and distribution of data locally available at each client on the predictive performance. We show that, federated learning is able to provide good predictive performance compared to other settings for most cases of label availability. Unlike synthetic data-based learning, which seems to highly depend on the type of training data, it consistently provides good predictive performance across different data sets. In addition, federated learning is able to perform well under different data distribution scenarios.