Edge-Aware Multi-Modal Intrusion Detection for Consumer-Centric Industrial Internet-of-Things Devices Using Deep Adaptive Fusion

The rapid proliferation of Consumer-Centric Industrial Internet-of-Things (CCIIoT) ecosystems has substantially expanded the attack surface for diverse cyber threats. Conventional intrusion detection systems (IDS) primarily rely on signature-based techniques, which exhibit limited effectiveness against zero-day and evolving attack patterns. To address these limitations, this study proposes a deep learning–driven IDS framework that concurrently integrates multiple modality-specific models. The proposed architecture employs a Gated Recurrent Unit (GRU) for host system log analysis, a Long Short-Term Memory (LSTM) network for network flow modeling, a 1D Transformer for packet-level traffic inspection, and a fine-tuned TinyBERT model for phishing email detection. The probabilistic outputs of these heterogeneous models are combined through a learnable adaptive fusion mechanism, enabling confidence-aware decision aggregation across modalities. Extensive experiments conducted on the UNSW-NB15, CIC-IDS2018, CTDAPD, and phishing email datasets demonstrate that the proposed framework outperforms conventional IDS approaches, achieving improved accuracy, F1-score, and ROC-AUC while significantly reducing false positive rates. The framework is explicitly designed as edge-aware through lightweight architectures edge-aware and late-fusion strategies, while detailed deployment benchmarking is reserved for future work.