Scalable Multi-Party Computation Protocols for Machine Learning in the Honest-Majority Setting

Watch a 13-minute conference presentation from USENIX Security '24 exploring innovative approaches to scalable multi-party computation protocols for machine learning in honest-majority settings. Discover how researchers from the University of Science and Technology of China, Shanghai Qi Zhi Institute, and Shanghai Jiao Tong University developed an efficient protocol using Damgaard-Nielsen methodology with Mersenne prime fields. Learn about the implementation of secure computing operations like truncation and comparison, and understand how the two-layer multiplication protocol from ATLAS was enhanced to optimize neural network operations. Examine impressive performance metrics showing the protocol's ability to complete oblivious inference of a 4-layer convolutional neural network with 63 parties in just 0.1 seconds in LAN settings and 4.6 seconds in WAN environments, demonstrating significant speed improvements over existing solutions like the Falcon protocol.