Explainable resource-Aware IoT security model via knowledge distillation and adaptive loss function optimization

• Developed RAID-KL, a teacher-student framework tailored for IoT security. • Validated RAID-KL on multiple data, achieving superior accuracy and efficiency. • RAID-KL achieves 11.3% and 64.33% reduction in CPU and memory usage, respectively. • RAID-KL compressed teacher model by 91.24% with an accuracy of approx. 99.75% • Applied SHAP to interpret feature contributions for IoT attack detection. Knowledge distillation (KD) is a pivotal model compression technique that enables the deployment of lightweight neural networks on resource-constrained Internet of Things (IoT) devices without sacrificing predictive performance. In the literature, KD has been widely applied to intrusion detection designs using the Kullback-Leibler (KL) divergence loss to align the distillation loss. However, KL divergence suffers from asymmetry and numerical instability when student predictions poorly match teacher distributions. The Jensen-Shannon (JS) divergence presents an alternative that allows both distributions to be treated equally relative to their average and ensures more balanced knowledge transfer process. The JS although symmetric and bounded, may converge more slowly due to its conservative nature. This study introduces RAID-KL , a resource-aware security model that leverages KD and a novel adaptive loss function combining hybrid KL and JS divergence. By integrating the benefits of both divergences, our method enhances the generalization and convergence of distilled models while maintaining low computational overhead, balanced knowledge sharing and improved numerical robustness. RAID-KL utilizes 1D Convolutional Neural Networks (1DCNNs) as the learning algorithm with teacher-student paradigm where a complex model serves as the teacher model, transferring its learned representations to a significantly smaller student network. RAID-KL is trained and evaluated in real-world network traffic datasets, including CICIoT2023, CICIoMT2024 and NIMSLABIoT2025, which include several IoT threats. RAID-KL demonstrates high performance in all applied metrics and low resource utilization during training and inference. To elucidate the critical relationships and feature contributions in model decisions, we integrate SHapley Additive exPlanations (SHAP) values for interpretability. Our findings reveal that the choice of loss function significantly impacts both performance and resource efficiency, with the hybrid KL-JS loss achieving superior trade-offs. Empirically, RAID-KL achieves 11.3% reduction in CPU usage and 64.33% reduction in memory usage during inference. Additionally, the RAID-KL model achieves 91.24% model compression over the teacher model while maintaining nearly the same classification accuracy. These insights highlight the robustness of RAID-KL framework, while providing a nuanced explainable index compared to the literature.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.