Neuro-Symbolic and Federated Learning for Privacy-Preserving Intrusion Detection

Abstract

The concept of artificial intelligence (AI) has established itself as a foundation of contemporary intrusion detection systems because it can examine the network traffic on a large scale and detect sophisticated attack patterns. Nevertheless, the issues of data privacy, interpretability, and class imbalance still remain as the hindrances to the efficiency of the available solutions. The paper performs a detailed experimental analysis of six AI-based intrusion detection models, Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), Artificial Neural Networks (ANN), Federated Learning (FL) and Neuro-Symbolic Artificial Intelligence (NSAI) on the CICIDS2017 dataset within a single framework of pre-processing and evaluation. The findings show that the tree-based models (RF and DT) exhibit the highest overall performance (0.9985), which is why they are appropriate in resource-limited and high-speed settings. The Artificial Neural Networks are suitable in the nonlinear pattern of traffic with a precision of about 0.9880. On competitive performance ([0.9838]), Federated Learning achieves competitive performance and obtains data privacy only with decentralized training, which is suitable in collaborative and privacy-conscious environments. Importantly, Neuro-Symbolic AI model has a high accuracy/interpretability/robustness balance, as it has better recall and F1-score with the minority and complex attack classes. These results suggest that there is no universally best model but instead the option of intrusion detection methodology depends on specific operational needs such as detection reliability, computing efficiency, and privacy limits. The research is practically applicable in choosing AI-based models of intrusion detectors in the actual world of cybersecurity.

DOI: https://doi.org/10.54633/2333-025-059-021