Code-Based Cryptography for Confidential Inference on FPGAs: An End-to-End Methodology

Confidential inference (CI) involves leveraging data encryption to safeguard privacy while allowing inference on encrypted data. Various cryptographic methods, such as homomorphic encryption or order-preserving encryption (OPE), are commonly employed for CI. In this work, we inspect the validity and efficiency of code-based cryptography for CI in FPGAs for the case of an ensemble of decision trees called the random forest (RF) machine learning (ML) model. FPGAs are an excellent platform for accelerating ML inference because of their low-latency performance, power efficiency, and high reconfigurability. However, creating hardware descriptions for encrypted ML models can pose challenges, especially for ML developers unfamiliar with hardware description languages. Thus, we propose an end-to-end methodology that includes high-level synthesis for ease of ML accelerator implementation on FPGAs. Additionally, we introduce variants of lightweight OPE tailored for CI in RFs. The successful and efficient implementation has been demonstrated using the Jet and MNIST datasets on the Xilinx Artix-7 FPGA.