PCPR: Plaintext Compression and Plaintext Reconstruction for Reducing Memory Consumption on Homomorphically Encrypted CNN

In the big data era, data privacy is a concern for everybody. Adopting homomorphic encryption is a promising way of preserving data privacy; however, it consumes large memory. Previously proposed lazy encoding encapsulates a vector into one data on demand, decreasing memory consumption. However, it results in 2.10–2.49× application latency increase in our experiment, compared to without lazy encoding. This paper proposes a novel technique called plaintext compression and plaintext reconstruction (PCPR), a lightweight pre-encoding and on-demand processing, which achieves almost the same memory consumption decrease as lazy encoding with a shorter latency. Our ideas are 1) dividing data into masks and corresponding scalars and 2) using lightweight operations instead of encoding. Experimental results show that PCPR achieves 1.16–2.03× shorter latency with 0.07–0.15 GiB larger memory consumption than lazy encoding, reducing memory consumption by 16.97–68.17 GiB compared to a method without lazy encoding and PCPR.