ABNN$^2$: Secure Two-party Arbitrary-Bitwidth Quantized Neural Network Predictions
TimeTuesday, July 12th4:18pm - 4:42pm PDT
Location3006, Level 3
Event Type
Research Manuscript
AI/ML Security/Privacy
DescriptionAs a prevalent business model, many technology companies provide Neural Network prediction services. However, the data is usually sensitive and confidential, which greatly arouses customers' concerns about data privacy.

In this work, we utilize the advantages of Quantized Neural Network (QNN) and MPC to present \textsc{ABNN}$^2$, a secure two-party framework for neural network predictions with arbitrary-bit-width quantization. We propose an efficient and novel matrix multiplication protocol based on 1-out-of-$N$ oblivious transfer extension protocols. Further, through the parallel scheme, we optimize the performance of the protocol and the experiments demonstrate that our protocols achieve better performance than previous work.