Entrada to Secure Graph Convolutional Networks

Graph convolutional networks (GCNs) are gaining popularity due to their powerful modelling capabilities. However, guaranteeing privacy is an issue when evaluating on inputs that contain users' sensitive information such as social interactions, financial transactions, etc. To address such privacy concerns, we design $\mathsf {Entrada}$, a framework for securely evaluating GCNs that relies on the technique of secure multiparty computation (MPC). For efficiency and accuracy reasons, $\mathsf {Entrada}$ builds over 4-party MPC framework of Tetrad (NDSS'22) and enhances it by providing necessary primitives. Moreover, $\mathsf {Entrada}$ leverages GraphSC paradigm of Araki et al. (CCS'21) to further enhance efficiency. To leverage the GraphSC paradigm, we (i) identify computations in GCN evaluation that benefit from GraphSC, and (ii) define $\mathsf {Scatter}\PLXtext{-}\mathsf {Gather}$ primitives of GraphSC specific to the identified computations. Further, as required for GraphSC, we design a secure and efficient shuffle protocol in the 4-party setting, which to the best of our knowledge, is done for the first time and may be of independent interest. Through extensive experiments, we showcase secure GCN's accuracy evaluated via $\mathsf {Entrada}$ is on par with its cleartext counterpart. We benchmark efficiency of $\mathsf {Entrada}$ with respect to the included primitives as well as the framework as a whole. Finally, we showcase $\mathsf {Entrada}$'s practicality by benchmarking GCN-based fraud detection applications.