This paper focuses on various types of attacks and errors in an activation process of Gigabitcapablepassive optical networks. The process sends messages via Physical Layer Operation Administrationand Maintenance header field inside the transmitted frame. An exemplar network communicationis captured by a special hardware-accelerated network interface card capable of processing opticalsignals from passive optical networks. The captured data is filtered of irrelevant parts and messagesand correctly formatted into a suitable shape for a neural network. The filtered data is divided intosmall sequences called time windows and analyzed using a recurrent neural network-based on Gatedrecurrent unit cells. A new neural network model is designed to classify sequences into several categories:additional message, missing message, error inside (noisy) message, and message order error.All of these categories represent a certain type of attack or error. The proposed model can distinguishmessage sequences into these categories with high accuracy resulting in revealing a possible attackeror drift from protocol recommendation.