Advances in Intrusion Detection Approach in Industry 4.0 Systems using Evolving Continual Deep Learning

Deep Learning (DL) detects intrusions in Industry 4.0 applications based on the Internet of Things (IoT) since it can grasp complex patterns from large datasets and identify irregularities. Current DL-based techniques for IoT-based Industry 4.0 suffer from catastrophic forgetting, computational inefficiency, security and privacy concerns, and false positive predictions. A new intrusion detection method for Industry 4.0 systems employing Evolving Continual DL (ECDL) addresses these issues. Model flexibility, accuracy, computational efficiency, security, and privacy are ECDL goals. ECDL includes local IoT model data gathering, pre-processing, feature engineering, and classification. In pre-processing, the problem of missing values and noisy data is addressed, and then it is normalized to facilitate the training process. In ECDL, each IoT model pre-processes and trains locally obtained data with advanced DL layers. The methodology is adaptable since the proposed DL layers are connected with Incremental Learning (IL), allowing for the integration of new IoT data and learning without the need for extensive re-training. The locally trained models are sent to the global model, where the Federated Learning (FL) technique is employed for aggregation. FL-based trained model disseminated to each local IoT model and SoftMax classifier for intrusion detection. In ECDL, FL with IL trains feature learning locally and adaptively in dynamic IoT contexts, resulting in complete efficiency. Implemented in real time using relevant IoT datasets, ECDL is assessed for accuracy, computational complexity, model flexibility, security, and privacy. Each IoT model uses Raspberry Pi, and the global server model uses Jetson Xavier. ECDL has 3.9% greater accuracy and 16.77% lower computing complexity than previous methods. The ECDL delivers effective security for Industry 4.0 design and implementation with model flexibility.