K-HashFed: Communication Efficient Federated Learning through Gradient Clustering and Hashing

Federated learning (FL) is a decentralized framework for collaborative model training while preserving data privacy. However, high communication overhead between clients and central server hinders FL scalability. This paper introduces K-HashFed, a novel gradient compression approach that integrates K-means clustering and locality-sensitive hashing (LSH) to minimize communication costs while preserving model accuracy. The proposed method clusters gradient updates and uses LSH to further merge similar centroids, to significantly reduce data transmission. Experiments conducted on benchmark datasets such as MNIST, Fashion MNIST, CIFAR-10, and CIFAR-100, demonstrates a significant reduction in data transmission—by orders of magnitude—compared to state-of-the-art methods, with minimal accuracy loss. The results validate K-HashFed as a communication-efficient, scalable solution for real-world FL systems.