Safeguarding IoT consumer devices: Deep learning with TinyML driven real-time anomaly detection for predictive maintenance

Internet of Things (IoT) security is paramount for enterprises, as it includes several strategies, techniques, actions, and protocols that aim to alleviate the high vulnerability of cutting-edge businesses. IoT consumer devices, from smart home appliances to wearable gadgets, have become ubiquitous daily, facilitating automation and seamless connectivity. However, ensuring their reliability and security presents a tremendous challenge. Anomaly detection methods offer a promising solution, especially those powered by TinyML (Machine Learning (ML) on Tiny Devices). These IoT devices can autonomously identify unusual behaviours or patterns that diverge from regular operation by leveraging the proficiencies of deep learning (DL) techniques enhanced for resource-constraint environments, like neural networks. Incorporating DL, anomaly detection, and TinyML allows real-time monitoring and proactive mitigation of malfunctions or security breaches in IoT devices. This advanced technology ensures improved reliability, privacy, and overall user experience in the dynamic landscape of connected devices, whether identifying irregular health data or detecting unauthorized access attempts on a smart door lock from the wearable fitness tracker. Therefore, this study develops a new Deep Learning technique to secure IoT consumer devices with TinyML Driven Real-time Anomaly Detection for Predictive Maintenance (DLTML-RTADPM). The DLTML-RTADPM technique aims to recognize and categorize the anomalies in IoT consumer devices. At the primary phase, the DLTML-RTADPM model normalizes the input data using Z-score normalization. In the DLTML-RTADPM method, the Fennec Fox Optimization Algorithm (FFA) is used for a high dimensionality reduction process where the optimal feature set is chosen. The DLTML-RTADPM technique implements gradient least mean squares with a bidirectional long short-term memory (GLMS-BiLSTM) approach for anomaly detection. To further improve the detection results of the DLTML-RTADPM technique, the Jaya optimization algorithm (JOA)-based hyperparameter tuning process is utilized. A series of simulations are performed on the benchmark dataset to ensure better detection outcomes of the DLTML-RTADPM model. The investigational validation of the DLTML-RTADPM method portrayed a superior accuracy value of 98.11% over other techniques.