End-to-End Secure IoT Device Attribution via Latent Space Regularized Autoencoding Networks

Abstract

The rapid expansion of smart homes has resulted in a wide range of interconnected IoT devices such as smart lights, thermostats, cameras, smoke detectors, and smart sockets each producing unique patterns of sensor readings and network behaviour. Accurate classification of these smart-home devices is crucial for ensuring household security, enabling intelligent automation, and supporting early anomaly detection. However, existing machine-learning methods such as K-Nearest Neighbours (KNN) and Random Forest suffer from several challenges, including sensitivity to imbalanced datasets, dependency on shallow hand-crafted features, and limited ability to distinguish between devices with similar behavioural patterns. This project proposes a hybrid IoT device classification framework that combines smart data balancing, comprehensive Exploratory Data Analysis (EDA), and a novel random feature extraction technique integrated with a deep autoencoder. The autoencoder generates a compact, noiseresistant latent feature representation that captures hidden relationships within the IoT sensor data, improving discriminative power and robustness. These enhanced features are then classified using Logistic Regression, selected for its computational efficiency, interpretability, and strong performance when applied to high-quality feature embeddings. The dataset includes multiple IoT device categories such as baby monitors, smart lights, motion sensors, security cameras, smoke detectors, smart sockets, thermostats, TVs, and smartwatches, each represented through multidimensional sensor readings. Smart data balancing techniques ensure fair representation of all device types, reducing bias and improving classification stability. The proposed framework significantly enhances classification accuracy compared to conventional methods while providing stronger resilience to noise, class imbalance, and overlapping feature distributions. By integrating automated feature learning through deep autoencoders with lightweight classification via Logistic Regression, the model offers a secure, scalable, and efficient solution for real-world IoT device identification tasks.