Host-based attack methods have grown increasingly sophisticated and stealthy, posing severe threats to network and information security. This has necessitated higher accuracy requirements for Host-based Intrusion Detection Systems (HIDS). While extensive research confirms the effectiveness of neural networks for intrusion detection, a critical challenge persists: traditional linear dimensionality reduction methods fail to capture the complex nonlinear features inherent in intrusion detection data (typically system call sequences). To address this limitation, we propose a deep autoencoder grounded in manifold learning by preserving local geometric structures during dimensionality reduction, unlike traditional autoencoders that focus solely on reconstruction loss. Our experiments leverage these manifold-informed reconstructions as inputs to generative detection baselines employing state-of-the-art clustering algorithms. The results demonstrate that the proposed approach excels at feature extraction and dimensionality reduction. Furthermore, we verify that filtering operational data against our generated baselines significantly reduces system overhead while maintaining detection accuracy.