The rapid expansion of memory and resource-constrained IoT devices has enormously increased vulnerability to cyber intrusions. Although deep learning-based intrusion detection systems (IDS) aim to improve intrusion precision, this improvement comes at the cost of increased inference latency and resource utilization. In this paper, we propose a lightweight CNN-GRU-based IDS model that utilizes mutual information-based feature selection to reduce input dimensionality, retaining the most informative features. The model is validated using four popular IoT security datasets: BoT-IoT, ToN-IoT, Edge-IIoTSeT and NSL-KDD. We employ SMOTE to reduce class imbalance in the training dataset for classifying both major and minor attacks. We achieve accuracies of 99.31%, 98.80%, 96.73%, and 94.20% on BoT-IoT, NSL-KDD, ToN-IoT and Edge-IIoTSeT respectively. Since inference latency is a critical requirement for resource-constrained IoT devices, the proposed model achieves inference times of 0.064 ms, 0.086 ms, 0.078 ms, and 0.073 ms on the respective datasets. These results demonstrate that the proposed IDS provides an effective balance between detection performance and computational efficiency for real-time IoT applications. In future we will focus on validating the proposed framework in real-world IoT deployment scenarios.