基於自適應剪枝的聯邦行人重識別個性化框架

Abstract

隨著智慧城市快速發展，行人重識別（Person Re-Identification, ReID）技術於公共安全與智慧監控領域的重要性日益凸顯。該技術的核心目標是從分散部署的異質攝影機網絡中，精準檢索並匹配同一行人的影像。 然而，實際應用面臨三大關鍵挑戰：首先，監控資料的敏感性所引發隱私問題，而多數研究透過聯邦式學習架構，將敏感性資料留在本地端進行分散式訓練，以避免原始數據的洩漏。然而，此方法亦衍生兩大技術瓶頸——其一為統計異質性（Statistical Heterogeneity）即跨場域數據分布的非獨立同分布（Non-IID）特性導致模型偏差；其二是傳輸延遲問題，因聯邦式學習反覆於邊緣與中心間交換模型所需的龐大通信負擔，而造成的顯著延遲問題。 為此，我們提出 FedKLPR——基於聯邦式學習架構所設計的輕量化行人重識別 (re-ID) 任務。該方法通過剪枝技術在確保模型準確率下降低於1% 的同時，分別在ResNet-50及ResNet-34的模型架構上降低35.01% 及30.58% 的傳輸負擔，以提升學習效率。總結來說，我們提出的FedKLPR在不犧牲隱私與準確率的前提下，顯著地降低通訊成本，同時增強模型的個性化能力與實用性，為基於聯邦式學習的行人重識別任務提供了一條可行的輕量化發展路徑。

As smart cities continue to evolve, the role of person re-identification (re-ID) has grown crucial in enabling effective public surveillance and enhancing urban safety measures. The fundamental challenge in person re-identification lies in developing robust cross-camera matching capabilities across diverse surveillance systems to reliably associate images of identical subjects. However, deployment on real-world faces three major challenges. First, the sensitive surveillance data raises significant privacy concerns. To address this, many studies adopt a federated learning framework, which enables decentralized training by keeping data on local devices, thereby mitigating privacy risks. Nevertheless, this approach introduces two critical technical bottlenecks. First, statistical heterogeneity, where the non-independent and identically distributed (non-IID) data across different domains leads to biased model performance. Second, communication latency, caused by the substantial transmission overhead incurred by frequent model exchanges between edge devices and the central server. To tackle these issues, we proposed FedKLPR, a lightweight person re-ID framework built upon the federated learning framework. By incorporating pruning techniques, FedKLPR reduces the communication cost by 35.01% on ResNet-50 and 30.58% on ResNet-34, while ensuring that the model accuracy drops by less than 1%. In summary, FedKLPR achieves a significant reduction in communication cost without sacrificing privacy or accuracy, and simultaneously enhances model personalization and practical applicability, offering a viable and efficient solution for federated person re-identification tasks.