Robust Classification of Wi-Fi Network Saturation Condition with the Sparse Data Observers (SDO) Outlier Detection Algorithm

Machine Learning (ML) algorithms have been increasingly applied to wireless communication problems in recent years. This paper addresses the problem of fair coexistence in the unlicensed spectrum of Wi-Fi with other co-located technologies (e.g., LTE) by identifying the saturation status of a Wi-Fi network. Such status is determined by applying ML algorithms to realtime and over-the-air collection of medium occupation statistics of the Wi-Fi frames. In particular, inter-frame spacing statistics of Wi-Fi frames are used to create a Random Forest (RF) model that can classify the Wi-Fi network saturation. As in other fields, the performance of the ML algorithms relies on the quality of the input data, which may not be guaranteed either unintentionally because of sensor faults or intentionally because of attacks aimed to confuse the ML algorithm. In both cases, there is the need to define robust ML algorithms, which can mitigate potential errors in the statistics collected from the network. This paper proposes an enhancement of the ML-based analysis process, which uses Outlier Detection (OD) algorithms to detect and filter out anomalies to sanitize the statistical data before the application of the ML algorithm. The OD-based approach is applied to a public Wi-Fi Saturation data set, where it is shown to mitigate significantly (even more than 4% classification improvement) the presence of perturbed data. In particular, the sparse data observers (SDO) outlier detection algorithm, recently introduced in the research literature, is shown to outperform other OD algorithms (e.g., iForest, LOF, SDOF, iNNE) considered in this study.

BALDINI Gianmarco;  SITHAMPARANATHAN Kandeepan; 

2025-03-31

IEEE

JRC140920

https://ieeexplore.ieee.org/document/11059652,    https://publications.jrc.ec.europa.eu/repository/handle/JRC140920,   

10.1109/IWCMC65282.2025.11059652 (online),