Designing In-Storage Computing for Low Latency and High Throughput Homomorphic Encrypted Execution

Data providers are concerned about data privacy when a third party handles their data. A typical example is cloud computing. In such cases, homomorphic encryption can be used to ensure data privacy. Although homomorphic encryption allows the execution of operations with encrypted data without decryption, its execution time is long. Previous studies have focused on the field-programmable gate array (FPGA)-based acceleration of homomorphic encryption. However, the overhead of data transfer between a host and an FPGA is a bottleneck owing to the limited bandwidth of a peripheral component interconnect express (PCIe). In-storage computing (ISC), i.e., using a computational storage device (CSD), can reduce the data-transfer overhead. In this study, a computational model coupled with a CSD is proposed to accelerate homomorphic operations. Simulations show that our proposed ISC model reduces latency by 71.2% and achieves 3. 5x throughput of a homomorphic encrypted convolutional neural network compared with 16 FPGA boards with 16 PCIe lanes and sufficient solid-state drives (SSDs), i.e., without CSDs.