FPGA Acceleration of Fully Homomorphic Encryption over the Torus

Fully Homomorphic Encryption over the Torus (TFHE) is a promising approach for secure computing in cloud servers to perform computations directly on encrypted data. However, TFHE has much higher computation complexity than its unencrypted counterpart. In this work, we propose an FPGA accelerator for TFHE computations. We illustrate the effects of an optimization called bootstrapping key unrolling on the tradeoff between performance of bootstrapping and FPGA resource consumption. We customize the data layout of TFHE ciphertext to optimize data access and improve data reuse. We parameterize the FPGA design for TFHE bootstrapping, which can be configured to achieve high performance for different user-specified security requirements and given FPGA resources. We implement our design on a state-of-the-art FPGA and compare it with existing results on CPUs. Our implementation for TFHE bootstrapping achieves 216x improvement in throughput and 16.5x improvement in latency compared with the software baseline on a state-of-the-art CPU server.