Threat Detection Framework for IoT Devices Using Adaptive Dilated Conv‐Gated Recurrent Units With Spatial and Temporal Attention‐Based Residual Autoencoder

Purpose Internet of Things (IoT) network is equipped with sensors, actuators, and act as communication capabilities. It enables device to collect, distribute, and share data using centralized techniques. This system offers real‐time monitoring, automation, and data‐driven intelligence, which are critical in many industries, such as healthcare, agriculture, smart cities, and manufacturing. A key factor in maximizing the stability of intercommunicated IoT systems and it has the capability to perform threat identification in sensor systems. Since, the proliferation of IoT has been improved in different sectors, their vulnerability also increased by generating various malicious activities and cyber‐attacks. Automatic identification of these threats should be performed to avoid the risks in the network and so, a deep learning‐aided threat detection model is proposed with neural systems to analyze anomalies in information and complex patterns to recognize potential security risks. Methods Here, a deep learning‐assisted threat detection approach is designed in the IoT environment for significantly recognizing the threat activities in an efficient manner. At first, the prescribed data are gathered in the benchmark Internet sources such as IoT Intrusion Detection, APA‐DDoS Dataset, Kddcup99.csv, RT‐IoT2022, and MQTT‐IoT‐IDS2020 then, the gathered data is given to the feature extraction model, where the deep features within the data are extracted by employing the spatial and temporal attention‐based residual autoencoder (ST‐ARAe) model. Further, the features extracted from the ST‐ARAe are given to the adaptive dilated conv‐GRU (ADC‐GRU) framework to detect the threat on IoT devices. Furthermore, the threat detection technique's performance is enhanced by tuning the network parameters on the ADC‐GRU by utilizing Random Function Enhanced Clouded Leopard Optimization (RFE‐CLO). Additionally, a set of tests is executed to reveal the effectiveness of the proposed approach. At last, the entire performance is evaluated and compared with several conventional methods to generate better outcomes in an efficient manner. Results Here, the proposed approach has attained 96% accuracy, 97% precision, 98% sensitivity, and 97% specificity in Dataset 3 when the system considers batch size as 48. Conclusion Thus, the proposed ADC‐GRU threat detection approach significantly proved that the superior detection performance than the conventional methods.