Federated Few-Shot Learning With Intelligent Transportation Cross-Regional Adaptation
Jiaming Pei, Marwan Omar, Maryam M. Al Dabel, Shahid Mumtaz, Wei Liu
Abstract
In Intelligent Transportation Systems (ITS), data scarcity and cross-regional distribution differences pose significant challenges to the development of robust machine learning models. To address these issues, we propose FFSLCRA (Federated Few-shot Learning with Cross-Regional Adaptation), a novel framework designed to handle data heterogeneity and scarcity in a federated learning setting. FFSLCRA integrates federated contrastive learning to align feature distributions across regions, region-specific data augmentation to enhance data diversity in underrepresented areas, and few-shot learning principles to enable robust adaptation to limited data scenarios. We provide theoretical guarantees for the proposed method, including convergence analysis and robustness proofs. Extensive experiments on two ITS benchmark datasets, BDD100K and GTSRB, demonstrate the superiority of FFSLCRA in accuracy, adaptability, and fairness compared to state-of-the-art methods. Our results validate FFSLCRA as an effective solution for federated learning in ITS applications with diverse and imbalanced data distributions.