A Game-Theoretic Federated Learning Approach for Ship Detection from Aerial Images

Detection and monitoring of ships in the images captured from satellites or aerial vehicles is a pivotal task in maritime security applications. Recent advancements in aerial communication and computer vision has enabled real-time collection of such images as well as development of robust and precise models for ship detection. However, conventional machine learning (ML) based models are prone to security and privacy issues as the real-time data captured through aerial imagery may be exposed during transfer or after storage in the cloud server. Furthermore, real-time decision making is a challenging task with conventional ML models due to the latency incurred while transmitting large amount of data from maritime aerial network to the cloud. To address the privacy and latency challenges, we propose a privacy-preserving game-theory based federated learning approach for ship detection in aerial images from maritime network. FL improves privacy by allowing raw data to reside at the edges/clients, and game theory helps in optimizing the parameter updates that are sent to the centralized server. Evaluation results prove the efficacy of the proposed model with a prediction accuracy of 96.01%, 92.96% reduction in time complexity and also 8.28% reduction in communication overhead.