Intrusion Detection in IIoT Using CNN-BiLSTM with Federated Learning and Attention Mechanism

Abstract

The increased utilization of Industrial Internet of Things (IIoT) networks has led to a greater need for effective and scalable intrusion detection systems. Yet, the issue of data privacy and security with high detection accuracy has been a concern. Centralized systems usually lack scalability since they are unable to decentralize privacy. In this paper, we seek to introduce a novel Hybrid Convolutional Neural Network - Bidirectional Long Short-Term Memory (CNN-Bi-LSTM) based IIoT Intrusion Detection framework using Federated Learning (FL). It involves the use of spatial feature detection capability provided by Convolutional Neural Networks (CNN) and temporal dependencies provided by Bi-LSTM. In order to prove the effectiveness of the proposed system, experiments were conducted on three benchmark IIoT datasets; namely, X-IIoTID, WUSTL-IIoT, and Edge-IIoTset. Data preprocessing involved normalization, feature encoding, and class balancing. In order to preserve the data privacy of users, FL method was utilized where models were trained locally and results aggregated centrally. Its performance was compared to the one obtained from centralized learning taking into account accuracy, precision, recall, F1-score, and AUC. The proposed hybrid CNN-Bi-LSTM model showed excellent intrusion detection ability with 97.8% accuracy and AUC higher than 0.95 on all datasets. The proposed approach improves an intrusion detection system for IIoT based on CNN-BiLSTM by using a multi-head attention mechanism. This method identifies important attributes from traffic data, enhancing attack detection efficiency while ensuring privacy with the help of federated learning. The empirical findings indicate the superior performance of the proposed approach.