The framework brings together several advanced techniques to deliver high accuracy with manageable computational demands. This balance has become increasingly important as smart homes become more connected and data-intensive.
Smart home systems rely on continuous streams of information from sensors, actuators, and networked appliances to monitor conditions, enhance security, and automate everyday tasks.
With IoT environments growing increasingly complex, they now require models that can interpret diverse and often noisy data while operating efficiently on real-world hardware.
Machine learning and deep learning have become central to this evolution, supporting more precise recognition of activity patterns and environmental changes. However, to ensure these systems remain both responsive and secure, we need lightweight, scalable models to handle these vast volumes of data produced.
Get all the details: Grab your PDF here!
Inside the LDEFF-SHADC Framework
The study presents a multi-stage pipeline designed to process and classify IoT device data reliably.
It begins with Linear Scaling Normalization (LSN), which places all features onto a consistent scale, typically between 0 and 1. This reduces the influence of outliers and stabilizes training across smart home signals.
Feature selection is handled using an Improved Snake Optimization (ISO) algorithm. Inspired by snake behavior, ISO identifies the most relevant features within high-dimensional IoT data. It reduces redundancy and computational load while helping prevent overfitting.
To model temporal and contextual relationships in device data, the researchers combined Gated Recurrent Units (GRU) with Multi-Head Attention (MHA). GRUs capture sequential patterns efficiently, while MHA focuses the model on the most informative elements across time.
Together, GRUs and MHA strengthen the system's ability to recognize both short- and long-term dependencies in device behavior.
Hyperparameter tuning is performed using the Improved Sparrow Search Algorithm (ISSA), another bio-inspired approach based on sparrow foraging behavior.
ISSA enhances the model’s optimization process by exploring the parameter space more effectively than conventional methods, reducing the likelihood of becoming stuck in local optima and improving overall convergence.
Benchmark Results
The authors evaluated LDEFF-SHADC using an IoT device recognition dataset containing 1,000 devices across 10 categories, ranging from security cameras and thermostats to wearable devices and motion sensors.
Across the full evaluation, the model achieved a maximum accuracy of 98.90 %, surpassing common baselines such as RNN-LSTM, MLP, XGBoost, and Attention-CNN.
Comparative experiments and ablation analysis demonstrated that each component of the framework (LSN, ISO, GRU-MHA, and ISSA) makes a meaningful contributions to the final performance, in both accuracy and computational efficiency.
Limitations and Future Directions
However, despite positive results, the authors note that the model’s performance comes with a few constraints.
As the framework has been trained on a specific dataset, its adaptation to more varied real-world IoT environments may be limited. Performance can be sensitive to noise in sensor readings, and computational demands may rise as datasets grow larger.
Future work will aim to improve adaptability across broader device ecosystems, incorporate additional data sources, and create lighter versions optimized for edge environments. Strengthening resistance to adversarial attacks and supporting continuous learning are also key priorities for improving long-term resilience.
Journal Reference
Alruwais, N. et al. (2025). A lightweight deep evidence fusion framework for smart home appliance detection and classification via internet of things devices. Scientific Reports, 15(1), 39399. DOI: 10.1038/s41598-025-99957-y
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.