Incentive-based Energy-efficient Federated Learning Aggregation for Intrusion Detection in IoT Sensor Network
C. Venkatesan, S Jeevanantham, B. Rebekka · 2023
At present, the wide range of critical applications encompassing healthcare and financial management are inhabited by wireless sensor networks (WSN). The foremost challenge associated with it is the resource constraint. Also, the WSN being itself deployed for critical applications, the adverse wireless security threats lead to head-on effect on the safety and operational efficiency. Hence in WSN, the intrusion detection implementation should corroborate the energy-efficient implementation for seamless operation. The inherent distributed characteristic of WSN mandates the federated learning to avoid the huge communication cost. This in turn preserves the security of critical sensor data used for training. However, the sensor nodes suffer from energy depletion due to local training and communication for federated model aggregation. Thus, in the proposed Incentive-based Energy-efficient Federated Learning Aggregation (IEE-FedAgg), the global model incremental training performance is considered for optimally scheduling the sensor nodes for learning. The aggregation number is further tuned based on the real-time performance exhibited. The simple multi-layer perceptron (MLP) and Long Short-Term Memory (LSTM) were considered for local training. Both the communication and computation energy consumption of the sensor nodes are considered for better analysis. The proposed IEE-FedAgg-LSTM model demonstrates the multi-class prediction accuracy of 99% and 98% for iid and non-iid data distribution respectively. Further, the proposed IEE-FedAgg mechanism attains the energy savings of 50% and 24% respectively for instrusion detection in iid and non-iid data distribution.