The growing deployment of IoT-enabled surveillance systems necessitates scalable, real-time, and privacy-preserving person re-identification (Re-ID) across distributed environments. Existing cross-domain methods, including biclustering collaborative learning (BCL), are limited by centralized training and static data assumptions. This paper proposes a novel Federated Multi-Modal Streaming Biclustering (FM²BCL) framework that enables adaptive cross-context identity learning. The approach integrates multi-modal identity disentanglement from RGB, thermal, depth, and beacon data, along with streaming temporal graph biclustering for continuous pseudo-label refinement. An uncertainty-aware conditional triplet loss is introduced to suppress noisy hard samples. To ensure privacy and scalability, a federated memory bank performs distributed prototype alignment without raw data sharing. Experimental results on standard benchmarks demonstrate significant improvements in accuracy, robustness, and temporal stability. The proposed framework establishes an effective solution for real-world IoT-based Re-ID systems, enabling distributed intelligence, continuous adaptation, and secure large-scale deployment.

Federated Learning,Domain Adaptation,person re-identification,Multi-Model Learning,Biclustering IoT Systems