The development of technologies has led to IoT devices entering critical areas such as healthcare and retail. This is leading to everyday devices being connected to the internet so that these devices can work together to improve their performance and support each other. This interconnectivity brings with it security vulnerabilities, the threat evolves dramatically and the attack surface on the network increases, leading to catastrophic losses. Despite decades of development, existing intrusion detection systems still face the challenge of improving detection accuracy, reducing the false alarm rate and detecting unknown attacks. Practical security approaches, notably ensemble deep learning have been introduced to tackle overfitting, enhance generalization, handle imbalanced Data, reduce False Positive and improve detection accuracy. In this paper, we present an ensemble deep learning architecture that integrates a voting policy into the structure of the model, facilitating the computation and learning of hierarchical patterns to detect mutated traffic and prevent spoofed traffic. The proposed system uses a deep learning autoencoder with XGBoost and Isolation Forest (IF) methods. Also, a stacking classifier is used to combine the predictions of XGBoost and the ensemble approach. The experimental analysis showed that this approach performed better compared to other intrusion detection systems (IDS) on different datasets, with a maximum accuracy of \(100\%\) on the BoT-IoT dataset. The performance of the system on the CICIDS2017 dataset in terms of accuracy is \(99.32\%\) .