R-SVDD: A decentralized radius-aware federated prototype object detection framework

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Federated Learning (FL) methods typically require Deep Neural Network (DNN) weight transfer from FL clients to an FL aggregator (master) for centralized DNN aggregation. However, in real-world settings such as smart-city sensing and emergency-response monitoring, weight transfer may be infeasible due to privacy regulations, operational policies, or unreliable connectivity. In this paper, we present R-SVDD, a fully decentralized FL framework that enables collaborative DNN training for object detection without neural parameter sharing. R-SVDD replaces weight exchange with a lightweight class-wise prototype learning mechanism based on Support Vector Data Description (SVDD), trained locally on each FL client using private object detection features. Collaboration is achieved by exchanging only SVDD class prototypes in a peer-to-peer topology, allowing event-specific appearance changes observed by one client to propagate as prototype-level adjustments that improve consistency across the federation. To prevent unreliable FL clients from corrupting this shared representation under extreme non-IID conditions, R-SVDD introduces a radius-aware reliability rule that uses the SVDD hypersphere radius as a confidence signal and suppresses outlier SVDD prototypes during aggregation. Neither DNN weights nor training data are exchanged between FL nodes whatsoever. R-SVDD is evaluated under multi-client configurations that simulate separate data sources across locations and disaster-related distribution shifts, demonstrating competitive object detection performance with significantly lower communication cost than state-of-the-art FL methods.