The rapid evolution of Smart Infrastructure (SI) on a global scale has revolutionized our daily lives, empowering us with unprecedented connectivity and convenience. However, this evolution has also exposed smart devices to increasingly sophisticated cyber-threats, endangering the integrity of entire smart networks. In response to these challenges, this paper proposes a novel approach utilizing Deep Learning (DL) models for multi-class threat detection in SI environments. Specifically, we introduce the Cu-GRULSTM model, which leverages CUDA-enabled Gated Recurrent Units (GRU) and Long Short-Term Memory (LSTM) architecture. Additionally, we employ the Cu-GRUDNN model for comparative analysis. Both models are trained and evaluated using the efficient and publicly available CICIDS2018 dataset. Our evaluation results demonstrate the superior performance of the proposed Cu-GRULSTM model, achieving an exceptional accuracy rate of 99.62% with a minimal False Alarms Rate (FAR) of 0.0003. This significant improvement over existing models underscores the efficacy of our approach in mitigating cyber-threats in smart infrastructure environments.