Efficient Decentralized Federated Singular Vector Decomposition

Explore a 19-minute conference talk from USENIX ATC '24 on Efficient Decentralized Federated Singular Vector Decomposition. Delve into the innovative Excalibur system, which addresses the challenges of federated singular value decomposition (SVD) in distributed applications. Learn how researchers from Hong Kong University of Science and Technology and Peking University developed a lightweight matrix protection method and a communication-efficient decentralized SVD workflow to overcome privacy and efficiency issues. Discover how Excalibur eliminates the need for external servers while achieving 3.1× to 6.0× faster performance than state-of-the-art server-aided methods on billion-scale data. Gain insights into the system's impressive throughput, which is over 23000× larger than current homomorphic encryption-based systems.