A Study on the Application of Personalized Federated Learning for Enhancing the Performance of Defense Surveillance Systems

국방 경계감시시스템의 성능향상을 위한 개인화 연합학습 기법 적용방안 연구

The military operates numerous CCTV systems for surveillance purposes and has recently begun integrating AI technology to improve operational efficiency. Current AI models applied to CCTV systems utilize pre-trained models based on large-scale training data; however, this approach presents several limitations in AI model performance management: restricted continuous and rapid model updates, limited optimization for operational environments, and challenges in model management due to varying AI models across manufacturers. Therefore research on AI learning methods that can overcome these limitations and achieve optimal performance in actual operational environments is necessary This study proposes personalized federated learning (PFL) techniques for AI model performance enhancement in CCTV systems based on an analysis of defense surveillance system operational environments. In the proposed approach, AI models in video processing servers with cross-silo characteristics apply identical AI models with PFL techniques based on model weight exchange for customized learning, while AI models in CCTV terminals with cross-device characteristics employ prototype-based PFL techniques. Experimental results confirm that the proposed PFL techniques demonstrate improved performance in classification accuracy and communication efficiency compared to other learning methods. Additionally, we identify environmental considerations for implementing personalized federated learning. The findings of this research are expected to contribute to AI application strategies and planning to maximize defense surveillance system performance in actual operational environments.