Abstract
<title>Abstract</title> <p>In recent years, the growth of Internet of Things (IoT) devices has increased due to their efficiency in automation, sharing of information across multiple devices and robust Intrusion Detection Systems (IDS). Traditional IDS approaches often struggle with high feature redundancy and class imbalance, which leads to inefficient performance in large-scale IoT deployments. To address these challenges, a Dual Feature Selection with Loss-Aware attention transformer for IDS (DFLA-IDS) is proposed. The proposed DFLA-IDS first employs a global feature selection stage with a Maximal Information Coefficient (MIC) and minimum Redundancy Maximum Relevance (mRMR) to eliminate irrelevant and redundant attributes. Next, a loss-aware attention mechanism embedded within a transformer model performs local feature refinement, which dynamically adjusts the weights of features at the instance level to ensure that classes are effectively represented by reducing data imbalance. The attention mechanism highlights the critical features; these weighted representations are passed through a SoftMax layer to identify intrusions in IoT networks. The proposed DFLA-IDS achieved better results in terms of accuracy (99.90%), precision (99.92%), recall (99.88%), and F1-score (99.85%) for UWSN-NB15 dataset when compared to deep bidirectional long short-term memory (Deep BiLSTM).</p>