Detection of invalid AIS messages using machine learning techniques

Abstract

Coastal nations monitor maritime activities in the interest of defence, security, and safety. This form of monitoring typically occurs at operations centers that visualize the maritime environment by creating a Recognized Maritime Picture (RMP) covering a particular area of interest. The creation of this picture changed drastically with the introduction of the Automatic Identification System (AIS). AIS messages are known to contain numerous types of errors and in April 2020 a unique error was found in the data stream. This error consisted of messages indicating the appearance of over 200 vessels in the North Atlantic taking part in a yacht race when in fact no race or physical ships existed. The following work explores the application of various machine learning (ML) techniques to help identify these types of fabricated AIS messages. Specific ML techniques were explored including: K-means clustering, Decision Tree (DT), Random Forest (RF), Feed-Forward Neural Networks (FNN), Support Vector Machines (SVM), and One-Class Support Vector Machines (One-SVM). The results showed that DT, RF, and FNN best identified the fabricated AIS messages with F1 scores greater than 93 percent on the test data.