B-CVAE: Bilinear conditional variational autoencoder with adaptive graph Laplacian regularization for IoT intrusion detection

The rapid proliferation of IoT has made resource-constrained devices prime targets for cyber intrusions, where severe sample space overlap in high-dimensional traffic severely hampers effective detection. We propose a novel manifold reconstruction framework that combines a bilinear-conditional variational autoencoder (B-CVAE) with adaptive batch-level graph Laplacian regularization. The B-CVAE captures high-order feature-condition interactions via bilinear mapping and residual fusion, while the Laplacian term dynamically preserves local manifold geometry to enhance intra-class compactness and inter-class separation. A lightweight classifier is applied to the reconstructed representations for final detection. Extensive experiments on NSL-KDD, RT-IoT2022, and CIC-IoT2023 datasets demonstrate mean accuracies reaching 99.6% with low standard deviations across multiple runs, outperforming recent autoencoder-based and hybrid research. The method offers a robust, efficient solution tailored to resource-limited IoT environments.