Adaptive cyber threat detection in internet of things environment using deep learning and metaheuristic optimization

Internet of Things (IoT)–based systems are increasingly becoming targets of sophisticated cyber threats due to their critical roles in scientific computation, defense, and large-scale data processing. Traditional intrusion detection systems often struggle to recognise the complicated patterns in IoT traffic, which means we need stronger and more flexible cybersecurity solutions. Current methods, despite their advancements in anomaly detection, struggle to efficiently learn high-dimensional, heterogeneous, and sequential features in IoT environments. This study suggests a new cybersecurity system that combines Convolutional Neural Networks (CNNs) to identify spatial features, Gated Recurrent Units (GRUs) to detect time-based anomalies, and eXtreme Gradient Boosting (XGBoost) for the final classification. To further optimise detection performance, the Prairie Dog Optimisation (PDO) algorithm is employed for automated hyperparameter tuning. The CIC-IDS 2018 and LANL Cybersecurity Datasets were modified to simulate IoT-specific threat patterns, including synthetic events such as MPI abuse and unauthorised job execution. Feature extraction was enhanced using CNNs and Symbolic Aggregate Approximation (SAX) for time series compression. GRUs were selected over traditional RNNs to address long-term dependencies and reduce training complexity. The escalation of cyber threats targeting IoT systems necessitates the evolution of cybersecurity mechanisms beyond traditional approaches. Rule-based systems and static models fail to The proposed CNN-GRU-XGBoost framework, enhanced with Prairie Dog optimisation, achieved an F1 score of 99.1% and a detection latency of 0.52 s, outperforming all baseline models by a notable margin. These results show that the framework can provide very accurate and quick intrusion detection, which is crucial for real-time cybersecurity in IoT settings.