IoT Anomaly Detection with Sound-Based Spectral Analysis and Lightweight Model
Hyeyoung Lee, Muhammad Nadeem · 2025
Internet of Things (IoT) networks have experienced an increase in popularity and adoption in recent times. This growing popularity of IoT network has led to attack by malicious groups. Moreover, these traffic attacks can directly affect the devices connected to the network, leading to even more severe consequences for classification. In this paper, we aim to detect anomalies in IoT network by integrating audio signal processing techniques using lightweight ResNet18-based 1D method. We perform the experiments on CICIoT2023, NSL-KDD, and IoTID20 datasets. Based on time series features with the stationarity in the pipeline as a medium of wire or wireless generates acoustic sound features. This feature extraction approach using audio-signal processing into spectral energy concentrated in the low-frequency area enhances deep learning (DL) classification performance in time-series data flow, which generates sound features. Experimental validation of the lightweight model is performed on Raspberry Pi (RPi) 4B, which demonstrates the effectiveness of our method for real-time anomaly detection across edge devices.