Towards neurosymbolic network intrusion detection

Although recent advancements in Deep Learning (DL) have significantly improved the performance of Network Intrusion Detection Systems (NIDS), many existing approaches still overlook model interpretability, limiting their effectiveness in real-world operational settings. To address this limitation, we propose a neurosymbolic framework for network intrusion detection that integrates the symbolic reasoning of a Bayesian Network with neural and statistical anomaly detection models. We evaluate our framework against a recent state-of-the-art method for open set recognition, a task that requires distinguishing known attacks from unknown ones. Results show that our framework achieves F1 scores comparable to the considered baseline, thus addressing the key dimension of explainability without sacrificing efficiency. Overall, our work represents a first step toward enhancing the interpretability of NIDS through the adoption of the neurosymbolic paradigm.