Unsupervised Anomaly Detection Using Bidirectional GRU Autoencoder Neural Network for PLOAM Message Sequence Analysis in GPON

This paper proposes an autoencoder neural network based on bidirectional gated recurrent unit layers used for anomaly detection in sequences of management protocol messages in gigabit-capable passive optical networks (GPONs). The autoencoder uses unsupervised learning, and the learning dataset is acquired from the real GPON network using a custom-made analyzer. The anomaly detection focuses on deviations in the management protocol in comparison to the baseline. It may indicate changes in the protocol itself caused by a different protocol implementation or a potential attack on the network. The capabilities of a trained autoencoder are evaluated on a generated dataset with various types of anomalies. The autoencoder reaches an average accuracy of 66% across all types of generated anomalies. However, the detection accuracy of sequences containing a high amount of random noise is 100%.