SAE-AM: Enhancing Network Intrusion Detection Using Sparse Autoencoders with Attention Modules

In response to the relentless evolution of cyber threats that continue to outpace traditional defence mechanisms, this study addresses key limitations in existing Intrusion Detection Systems (IDS), particularly those related to high dimensionality and computational inefficiency. We propose a novel framework, sparse autoencoders with attention modules (SAE-AM), which integrates SAE with both channel and positional attention mechanisms to enhance feature representation and optimise resource utilisation. While SAE effectively perform dimensionality reduction, the attention modules capture global dependencies across diverse input features. Leveraging a deep learning model, specifically a multi-layer perceptron classifier – our framework efficiently classifies normal and attack samples. Extensive evaluations on benchmark network intrusion datasets, CICIDS2017, NSL-KDD, and UNSW-NB15, demonstrate the robustness and superior performance of the proposed method, achieving 99% accuracy and minimal false alarm rates across all datasets. SAE-AM makes significant strides in overcoming the core challenges of NIDS by reducing dimensionality and improving computational efficiency. This novel appro