Deep-learning-based Network Intrusion Detection Systems (NIDS) are known to struggle with attack classes that never appeared in their training data. This limitation leaves them vulnerable to zero-day threats, which by definition are novel and unseen. The property that determines how well a model handles such unseen attacks is called transferability, and it is the focus of a new arXiv paper.
The authors propose an adaptive and proximal federated optimization approach to improve this transferability. Rather than relying on a single centralized training set, the method leverages federated learning principles, where models are trained across distributed data sources while adapting optimization steps. The proximal term helps maintain stability during training, which appears to help the model generalize beyond the specific attack classes it has seen.
While the abstract does not provide experimental results, the work addresses a well-known weakness in NIDS: the gap between training on known attack patterns and performing reliably in the wild. If effective, the approach could make intrusion detection systems more resilient to novel threats without requiring constant retraining on new attack signatures.