CompGC: Efficient Offline/Online Semi-Honest Two-Party Computation

We introduce a new technique, component-based garbled circuits, for increasing the efficiency of secure two party computation in the offline/online semi-honest setting. We observe that real-world functions are generally constructed in a modular way, comprising many standard components for common tasks like arithmetic or cryptographic operations. Our technique allows circuits for these common tasks to be garbled and shared during an offline phase; once the function to compute is specified, these pre-shared components can be chained together to create a larger garbled circuit. We stress that we do not assume that the function is known during the offline phase only that it uses some common, predictable components. We give an implementation, CompGC, of this technique and measure the efficiency gains for various computations. We compare first to standard garbled circuit-based secure two party computation, where we find that our technique results in roughly an order of magnitude performance improvement. We then consider a set of machine learning classification computations previously studied by Bost et al. (NDSS 2015) that do not use garbled circuits. We find that our component based technique can improve online performance in most cases, including an order of magnitude improvement for decision tree classification.