Urban Anomaly Detection by processing Mobile Traffic Traces with LSTM Neural Networks

Detecting urban anomalies is of upmost importance for public order management, since they can pose serious risks to public safety if not timely handled. However, monitoring large metropolitan areas requires complex systems that can potentially lead to elevated costs. In this paper, we discuss the opportunity of exploiting the mobile network as a supplementary sensing platform for detecting urban anomalies. To favour the reliable and low latency anomaly recognition, we rely on a Multi-access Edge Computing (MEC) architecture, which enables a deep and detailed mobile traffic characterization almost in real-time and allows for a performance-responsive service, that is crucial in our problem. We focus on urban anomaly detection, by monitoring known events that gather a high concentration of people. The mobile network information is collected from LTE Physical Downlink Control Channel (PDCCH), which contains the radio scheduling information and has the benefit of being unencrypted and fine-grained, since the messages are exchanged every LTE subframe of 1 ms. To this purpose, we design an anomaly detection system based on Long Short-Term Memory (LSTM) Neural Networks, to deal with sequential and recurrent inputs. We demonstrate that a stacked LSTM architecture is able to identify traffic anomalies provoked by a rapid growth in the number of users, when a crowded event takes place nearby the monitored area. The numerical results show that the proposed algorithm reaches an F-score = 1 and overcomes the performance of other state-of-the-art benchmarks.