Improving detection of scanning attacks on heterogeneous networks with Federated Learning

Scanning attacks are the first step in the attempt to compromise the security of systems. Machine learning (ML) has been used for network intrusion detection systems (NIDS) to protect systems by learning misbehavior based on network traffic. This paper demonstrates that Federated Learning (FL) is a promising approach to achieve better detection performance than traditional local training and inference on distributed agents. Also, this FL approach brings privacy, efficiency, and it is suitable for distributed ML-based NIDS solutions. We present a horizontal FL setup using Logistic Regression with FedAvg strategy applied to 13 agents (data silos) capable of providing an iterative process of constant learning improvement. Our results indicate a more stable learning process when observed the F1-score average, whereas the traditional NIDS approach (local trained models) present lesser performance and bigger variability to classify scanning and benign traffic. We tested our model performance on the TONIoT dataset containing network traffic from a virtualized heterogeneous network composed of cloud, fog, and edge layers.