Traffic Anomaly Detection Using Deep Semi-Supervised Learning at the Mobile Edge
In this paper, we design an Anomaly Detection (AD) framework for mobile data traffic, capable of identifying different types of anomalous events generated by flash crowds in metropolitan areas. We state the problem using a semi-supervised approach and exploit the great performance of different Recur...
| Autores: | , , |
|---|---|
| Formato: | artículo |
| Estado: | Versión publicada |
| Fecha de publicación: | 2022 |
| País: | España |
| Recursos: | Centre Tecnològic de Telecomunicacions de Catalunya (CTTC) |
| Repositorio: | r-CTTC. Repositorio Institucional Producción Científica del Centre Tecnològic de Telecomunicacions de Catalunya (CTTC) |
| OAI Identifier: | oai:cttc.fundanetsuite.com:p7636 |
| Acesso em linha: | https://cttc.fundanetsuite.com/Publicaciones/ProdCientif/PublicacionFrw.aspx?id=7636 |
| Access Level: | acceso abierto |
| Palavra-chave: | Deep learning Urban areas Real-time systems Proposals Data models Process control Long Term Evolution Anomaly detection mobile networks traffic modeling smart cities remote sensing edge computing distributed learning |
| Resumo: | In this paper, we design an Anomaly Detection (AD) framework for mobile data traffic, capable of identifying different types of anomalous events generated by flash crowds in metropolitan areas. We state the problem using a semi-supervised approach and exploit the great performance of different Recurrent Neural Network (RNN) models to learn the temporal context of input sequences. Our proposal processes real traffic traces from the unencrypted LTE Physical Downlink Control Channel (PDCCH) of an operative network, gathered during an extensive measurement campaign in two major cities in Spain. The AD framework is designed to perform: i) a-posteriori analysis to understand users' behavior and urban environment variations; ii) real-time analysis to automatically and on-the-fly alert urban anomalies; and iii) estimation of the duration of the periods identified as anomalous. Numerical results show the higher performance of our AD framework compared to classic AD algorithms and confirm that the proposed framework predicts anomalous behaviours with high accuracy and regardless of their cause. |
|---|