Federated Learning for Anomaly Detection in Maritime Movement Data

TL;DR

The paper presents M³fed, a Flower-based federated extension of the M³ movement model for maritime anomaly detection, addressing privacy and bandwidth constraints in distributed AIS data. It demonstrates that M³fed achieves anomaly detection results that are comparable to the centralized M³ while dramatically reducing data transmission compared to centralized training. Through experiments on AISDK data with three overlapping coverage areas, the study analyzes anomaly events and quantifies communication savings, highlighting practical benefits in low-bandwidth maritime environments. The work also discusses limitations, such as ground-truth absence, and outlines future directions for model tuning, compression, and deployment on mobile vessels.

Abstract

This paper introduces M3fed, a novel solution for federated learning of movement anomaly detection models. This innovation has the potential to improve data privacy and reduce communication costs in machine learning for movement anomaly detection. We present the novel federated learning (FL) strategies employed to train M3fed, perform an example experiment with maritime AIS data, and evaluate the results with respect to communication costs and FL model quality by comparing classic centralized M3 and the new federated M3fed.

Figures (7)

• Figure 1: Locations of AISDK records (white points) and the three fictitious AIS antennas (orange stars) and their modelled overlapping coverage areas (blue circles) in the area around Gothenburg.
• Figure 2: Prototypes learned by the three clients in round 1: red = client 0; blue = client 1; yellow = client 2. The size of the markers indicates the number of AIS records involved in training the respective prototype.
• Figure 3: M³ (red) and M³fed (blue) prototypes in comparison. The size of the markers indicates the number of AIS records involved in training the respective prototype.
• Figure 4: Anomaly assessment (comparison of anomalies from M³ and M³fed for ship type cargo): TP (blue), TN (gray), FP (pink), FN (red).
• Figure 5: Anomaly events of vessel 265501910: M³fed events (black) on top of M³ events (red) illustrating the m:n relationships of anomaly events and the overlaps between events.