Federated Hash Learning Boosts Privacy and Efficiency

This paper introduces a novel approach called Federated Hash Projected Latent Factor (FHPLF) learning, designed to address the challenges of efficiency and privacy in decentralized machine learning. Traditional Hash Learning (HL) often requires centralizing sensitive user data, while most Federated Learning (FL) methods incur high communication costs and privacy risks by transmitting large real-valued gradients. FHPLF aims to integrate the benefits of both. FHPLF's core innovations include replacing real-valued gradient matrices with compact binary gradient-like matrices, which drastically cuts down computation, storage, and communication overhead while bolstering privacy. It also employs Projected Hamming Distance for similarity modeling, enhancing the representational capacity of binary codes. Furthermore, a Secure Binary Gradient Reassembly and Privacy-Enhanced Upload (SBG-PEU) strategy is proposed to minimize user interaction leakage during data transmission. Extensive experiments across four datasets demonstrate FHPLF's superior performance over state-of-the-art HL and FL methods, achieving an optimal balance between accuracy, efficiency, and privacy.