Transformer-Based Object Detection with Privacy-Preserving Federated Learning and Collaborative Edge Computing

Abstract

This research introduces a pioneering hybrid framework designed to address the dual challenges of data privacy and
resource-limited deployment in Transformer-based object detection systems. The proposed system leverages the
combined strengths of Federated Learning (FL) for decentralized, privacy-preserving training and Collaborative Edge
Computing for efficient distributed inference. By applying this framework to three leading object detection models—
DETR, RT-DETR, and Faster R-CNN—we perform an exhaustive evaluation, demonstrating that our approach delivers
a notable trade-off between performance, privacy, and scalability. Our results show 286 mAP for DETR, 380 mAP for
Faster R-CNN, and 546 mAP for RT-DETR on COCO validation sets, while simultaneously achieving 95.3% reduction
in communication overhead relative to centralized training. Notably, Strong Differential Privacy (ε = 1.0) offers 90%
privacy protection at the cost of only a 3% accuracy drop. The framework is scalable up to 10 distributed clients,
ensuring robust privacy-preserving performance in real-world settings. Additionally, our approach offers superior
fault tolerance through component-level decomposition in inference pipelines, which facilitates continuous operation despite partial system failures. The scalable hybrid architecture is detailed through a three-layer design comprising a Coordination Layer, Collaboration Layer, and Edge Layer, fostering a seamless integration of FL and edge computing for diverse object detection applications. This work makes significant contributions to the privacy and efficiency of AI systems, with practical implications for industries dealing with sensitive data. Our framework provides a blueprint for deploying high-performance object detection models in privacy-sensitive, resource-constrained environments, marking a significant step towards distributed, privacy-aware AI systems that combine cutting-edge performance with real-world deployability.