A Study of Matrix Multiply by Homomorphic Encryption for Utilizing in Deep Learning Frameworks

2019 Fiscal Year Final Research Report

• Project/Area Number: 18K19786
• Research Category: Grant-in-Aid for Challenging Research (Exploratory)
• Allocation Type: Multi-year Fund
• Review Section: Medium-sized Section 60:Information science, computer engineering, and related fields
• Research Institution: Waseda University
• Principal Investigator: Kimura Keiji 早稲田大学, 理工学術院, 教授 (50318771)
• Co-Investigator(Kenkyū-buntansha): 和田 康孝 明星大学, 情報学部, 准教授 (40434310)
• Project Period (FY): 2018-06-29 – 2020-03-31
• Keywords: 秘密計算 / 準同型暗号 / 高速化 / マルチコア / アクセラレータ / FPGA

Outline of Final Research Achievements

This research aims at accelerating matrix-multiply in homomorphic encryption toward utilizing it in deep learning frameworks. Through the research, we obtained 5.53x and 3.73x speedups in maximum for two important computational parts in the target encrypted matrix-multiply process. In addition, we have developed a data transfer unit, which can quickly provide required data to accelerator hardware units. We also investigated and evaluated the relationship between the precision of computations and calculation time to reduce the calculation cost while keeping the appropriate precision. As a result, we obtained 8 points accuracy improvement and 54% speedup for image recognition at the same time by parallel inference with eight smaller neural networks.

Free Research Field

計算機システム

Academic Significance and Societal Importance of the Research Achievements

準同型暗号により暗号化したまま計算可能なことで、秘密を保ったままクラウドなどの第三者環境にデータを提供し安全に計算処理を行うことができるようになったが、その計算コストが極めて大きいことが問題となっていた。本研究により得られた成果により、準同型暗号による行列積の処理を高速化可能となる。行列積は深層学習処理の主たる計算要素であるため、秘密を保ったままにしてクラウドで深層学習処理（主に推論処理を想定）を行い、結果を安全に利用者に返すことが可能となる。