Client aware adaptive federated learning using UCB-based reinforcement for people re-identification

Abstract

People re-identification enables locating and identifying individuals across different camera views in surveillance environments. The surveillance data contains personally identifiable information such as facial images, behavioral patterns, and location data, which can be used for malicious purposes such as identity theft, stalking, or discrimination. This raises serious ethical and privacy concerns. The communication overhead of transporting a large number of data needed to train a global model and the diverse nature of the data from different sources are serious limitations facing the development of people re-identification technologies. We address these challenges by proposing a novel three-step federated learning framework. First, we investigate the impact of data augmentation techniques on the model generalizability and explore the effectiveness of different backbone networks. Second, we use reinforcement learning-based Upper Confidence Bounds (UCB) as a client-selection strategy in the federated round that dynamically chooses devices similar to the current model state, ensuring the model is updated with relevant data and enables faster convergence. Finally, we introduce a feature-level attention mechanism focusing on discriminative features for re-identification. Extensive experiments were conducted on nine benchmark re-ID datasets. The proposed framework outperformed the federated re-ID baseline by 10% in rank-1 accuracy and achieved results comparable to the centralized approach, with a difference of 2%. This improvement over the previous state-of-the-art establishes a new benchmark for federated re-identification.