Metric Annealed Federated Learning (MAFL) - an efficient and hierarchical approach towards NIDS for IoT edge devices

Abstract

The proliferation of Internet of Things (IoT) devices has created unprecedented security challenges for network intrusion detection systems. Starting from simple implementation, the usage of IoT has expanded in many industries to increase their robustness. Unfortunately, the drawbacks of such distribution are also emerging at an alarming rate. Gradually attackers are developing new methods to gain unauthorized access to an IoT network. Over the past few years, the frequency of attacks like Distributed Denial of Service (DDoS) has significantly risen. The importance of protecting network and data has risen sharply. Current state-of-the-art approaches remain insufficient to provide a comprehensive solution in this field. Numerous studies have been attempted using deep learning architectures and supervised machine learning methodologies for binary and multiclass classification tasks. However, these efforts have been consistently hindered by class imbalance issues within the available datasets, resulting in suboptimal model performance and limiting the practical applicability of these approaches. Moreover, literature on compromised node detection are faintly in number and light-weight machine learning models had a false positive rate of around 30 percent overall. In this meticulous study, Deep Learning models perform with high accuracy (99%) in binary classification although they struggle with multi classifications. Additionally, ubiquitous data-driven approaches are parameter heavy and executed in a controlled environment. Given the scarcity of real-world dataset the study requires distributed simulation. Hence, a new Metric Annealed (MA) Federated Learning is proposed with Flower Framework to represent real-life scenarios and prioritize security. To address the resource constraints, hierarchical CNN-BiLSTM client model is introduced with an adaptive probabilistic client selected MA averaging aggregation strategy. The hierarchical model offers an optimized number of parameters enabling the simultaneous training of binary and multi-classification heads. The proposed model achieved an accuracy of 98.8% in binary classification and 81.1% in multiclass classification while having low false positive rate, achieving efficient performance and resource utilization compared to state-of-the-art DL and FL models in precision, recall, and F1-score. Hyperparameter tuning through systematic tuning identified optimal temperature decay and weight parameters for energybased evaluation . Comparative results demonstrate that the MAFL model offers superior scalability, privacy preservation, and detection performance compared to centralized IDS approaches. The research establishes Metric Annealed Federated Learning (MAFL) as a viable and efficient framework for secure intrusion detection in distributed IoT networks.