Improved Stacked Autoencoder Combined with LSTM-CA for Intrusion Detection Model

To address imbalanced data challenges in intrusion detection, we propose SA-LCA, integrating an improved stacked autoencoder with LSTM-CNN-Attention. Preprocessing uses one-hot encoding and normalization to handle data discreteness. The enhanced autoencoder with Dropout/Attention reduces dimensionality while retaining discriminative features. LSTM captures temporal dependencies, and multi-layer CNN-Attention extracts spatial-temporal features. EQLv2 loss dynamically balances gradients to improve minority class detection. Experiments on NSL-KDD/UNSW-NB15 achieve 99.55%/84.53% accuracy, outperforming state-of-the-art methods.