The growing volume and sophistication of cyberattacks have driven research into effective network intrusion detection systems (NIDS). Anomaly-based network intrusion detection approaches have become increasingly prominent due to their effectiveness in identifying unknown and sophisticated attacks. However, many existing methods rely on centralized architectures, raising significant privacy concerns due to the need for aggregating sensitive data in one location. Additionally, a common assumption in these methods is the availability of only normal data, which does not reflect real-world scenarios where limited attack samples are often available. In this paper, we propose, PrivFlow-NIDS, a semi-supervised, decentralized approach that combines Normalizing Flows (NFs) and Federated Learning (FL). The NF model is leveraged for data augmentation enriching scarce anomalous samples. By learning the underlying distribution of normal traffic, NF generates pseudo-attacks from the distribution’s tails. Meanwhile, FL ensures privacy-preserving collaborative model training across distributed NIDS. We evaluate PrivFlow-NIDS using three well-known datasets, USTC-TFC2016, CIC-IDS2017, and UNSW-NB15. The results demonstrate that, with limited anomalous samples, PrivFlow-NIDS achieves high accuracy, detection rate, and macro F1-score, coupled with a low false alarm rate (99.91%, 87.59%, and 98.28% F1-score for USTC-TFC2016, CIC-IDS2017, and UNSW-NB15 respectively). Moreover, PrivFlow-NIDS performs on par with centralized settings, with differences of less than 0.1% in most cases. These findings highlight that integrating NFs and FL within IDS offers a promising, privacy-preserving solution for large-scale network intrusion detection.