Mobile Ad-hoc Networks (MANETs) face unique security challenges due to their dynamic topology, distributed nature, and resource constraints. While deep learning approaches have shown promise in intrusion detection, their computational overhead often conflicts with MANETs’ resource limitations. This paper presents a novel hybrid architecture combining Variational Autoencoders (VAE) and one-dimensional Convolutional Neural Networks (1D-CNN) for efficient intrusion detection in MANETs. Our approach leverages VAE for dimensionality reduction while maintaining detection accuracy through 1D-CNN classification. The proposed system first utilizes a 1D-CNN model for baseline intrusion detection, reaching an accuracy of 95.79% across multiple attack types. Then, we incorporate a VAE for feature space reduction and reduce the feature dimensionality by 16.67%, while the detection accuracy is maintained at a robust value of 91.29%. Experimental results confirm that the VAE-1D-CNN hybrid architecture can provide a desirable trade-off between resource efficiency and detection performance. In this paper, VAE reduces a great deal of feature space with only an accuracy loss of 4.5%. Then our proposed approach has been validated on comprehensive performance metrics: precision is 91.30%, recall is 91.29%, and F1-score is 91.27%. Confusion matrix analysis reveals that the hybrid model maintains reliable detection capabilities with 2562 true positives and 2038 true negatives, despite a modest increase in false positives. The model's training behavior shows stable convergence and consistent validation performance, indicating robust generalization capabilities. The contribution of this research is in the arena of resource-aware intrusion detection systems developed with MANET in mind. In essence, it provides a pragmatic solution to achieving both security effectiveness and computational efficiency.