Work

Enhancing data privacy in maritime operations with federated learning : a YOLOv8 object detection approach

Year: 2025
Type: dissertation
Abstract: The maritime industry is undergoing a transformation, driven by the adoption of artificial intelligence (AI). There are some concerns about the privacy and security of sensitive data. This thesis addresses the challenge by proposing a Federated Learning (FL) framework specifically adapted for maritime environments to enable collaborative model training without the need of sharing raw data. The approach uses the YOLOv8 object detection model and the TDSS-G1 dataset, which has real and synthetic maritime images. Each vessel trains a local YOLOv8 model using its own dataset and only shares updated model weights with a central server. These updates are then aggregated using the Federated Averaging (FedAvg) strategy to create a global model. In this thesis, the FL framework is compared to centralized learning systems, with performance indicators including mean average precision (mAP), recall, and training loss. The findings indicate that FL has the potential to be a scalable and privacy-friendly solution for real-time object detection in maritime operations. This thesis contribute to the development of secure, decentralized AI applications for use in ship-based systems, enhancing both safety and situational awareness at sea.
Author: Vo, Thao
Language: English
Open Access status: green
OpenAlex ID: W7078793932
Cites: 24
Cited by:
Related to: 10