Edge computing model for face detection and recognition based on federated learning for university

Face recognition has gained significant attention over the past three decades due to its pivotal role in image analysis and authentication systems. This study explored Federated Learning (FL) techniques to optimize face recognition training while addressing limitations in centralized data processing and ensuring compliance with the Data Privacy Act. Centralized approaches often face challenges such as high computational loads, increased latency, and reduced data privacy. Federated Learning, which is a decentralized method introduced by Google, preserves user privacy by training models locally and securely aggregating updates, minimizing communication costs and privacy risks. The study focused on the modeling of distributed computing and federated learning techniques for facial recognition in various multi-campus universities. Key contributions include preprocessing techniques, such as face detection via MTCNN, image resizing, and embedding extraction using FaceNet to generate robust datasets. A custom classifier model featuring Batch Normalization, Dropout layers, and ReLU activation was designed to enhance training stability and prevent overfitting. The result demonstrated the seamless training of local models and their aggregation into a global model, achieving scalability and reduced latency. Although global models showed slightly lower accuracy due to parameter adjustments, they exhibited robustness across diverse datasets, validating the efficacy of FL in distributed face recognition tasks.