ADAPTIVE MODEL OF INTELLIGENT TRAFFIC FILTERING IN IOT NETWORKS BASED ON TINYML AUTOENCODERS

Abstract

This paper presents and scientifically substantiates an adaptive model of intelligent traffic filtering for energy-efficient Internet of Things (IoT) networks. The proposed approach is based on the deployment of compact TinyML neural models using an autoencoder architecture directly in the firmware of ESP32 microcontrollers. Unlike conventional IoT systems that continuously transmit raw telemetry to cloud services, the developed model performs primary analysis locally at the edge node and supports an event-driven communication paradigm. Sensor data are segmented into fixed-length windows, normalized using Min-Max scaling, and processed by a lightweight autoencoder with a 10-8-4-8-10 architecture trained in an unsupervised manner to reconstruct normal signal patterns. Anomalies are detected by evaluating the mean squared reconstruction error (MSE) and comparing it with an adaptive threshold obtained during calibration. To ensure efficient execution on resource-constrained hardware, the trained model was converted using full integer quantization to Int8 format and deployed with TensorFlow Lite Micro. Experimental results confirmed the feasibility of running inference on ESP32 with a Tensor Arena of about 30 KB and an average inference time of 14 ms. The proposed method transforms the IoT node from a passive data transmitter into an intelligent edge filter that activates data transmission to Azure IoT Hub only when statistically significant deviations from normal operation are detected. Such selective communication reduces redundant outbound traffic by 85–95%, decreases channel load, minimizes radio module active time, and significantly extends battery-powered node lifetime. The developed model demonstrates high practical value for scalable industrial and monitoring IoT systems requiring low-power operation, anomaly awareness, and efficient cloud interaction.