A Dual-Head Gated Multi-Task and Uncertainty-Aware Deep Learning Framework for Intrusion Detection in Internet of Things Environments

Abstract

The rapid proliferation of Internet of Things (IoT) devices has significantly expanded the attack surface of modern cyber-physical infrastructures, exposing interconnected systems to a diverse spectrum of cyber threats. Conventional intrusion detection systems (IDSs) often suffer from limited generalization capability, inability to distinguish between anomaly detection and attack categorization simultaneously, and a lack of confidence estimation mechanisms for security-critical decision making. To address these challenges, this paper proposes DGMU-Net (Dual-Head Gated Multi-Task and Uncertainty-Aware Network), a novel deep learning framework designed for intelligent intrusion detection and classification in IoT environments. The proposed architecture integrates four key components: a residual deep neural decision module for high-level feature reasoning, a feature gating layer for adaptive feature selection, a dual-head multi-task output structure for simultaneous anomaly detection and attack classification, and an uncertainty-aware decision module that estimates prediction reliability through entropy-based uncertainty quantification. Unlike conventional IDS architectures that produce deterministic outputs, DGMU-Net introduces uncertainty-aware reasoning capable of identifying ambiguous attack instances requiring additional inspection. The residual learning mechanism enhances feature representation while mitigating gradient degradation, whereas the gating module suppresses redundant and noisy information before classification. Multi-task learning enables the network to jointly learn binary intrusion detection and multi-class attack categorization, improving generalization performance and feature sharing across tasks. Experimental evaluations conducted on benchmark IoT intrusion datasets demonstrate that the proposed framework achieves superior detection accuracy, improved attack classification performance, enhanced robustness against class imbalance, and reliable uncertainty estimation. The obtained results indicate that DGMU-Net constitutes a promising next-generation intelligent cybersecurity framework capable of supporting trustworthy and explainable intrusion detection for large-scale IoT ecosystems.