Acceleration of HE-Transformer with bit reduced SEAL and HEXL

• Item type: SIG Technical Reports(1)
• 公開日: 2022-07-20
• タイトル: Acceleration of HE-Transformer with bit reduced SEAL and HEXL
• タイトル / 言語: en
• タイトル: Acceleration of HE-Transformer with bit reduced SEAL and HEXL
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: アクセラレータ
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者所属: Department of Computer Science and Engineering, Waseda University
• 著者所属: Department of Computer Science and Engineering, Waseda University
• 著者所属: Department of Computer Science and Engineering, Waseda University
• 著者所属(英): en
• 著者所属(英): Department of Computer Science and Engineering, Waseda University
• 著者所属(英): en
• 著者所属(英): Department of Computer Science and Engineering, Waseda University
• 著者所属(英): en
• 著者所属(英): Department of Computer Science and Engineering, Waseda University
• 著者名: Xinyi, LI; Masaki, Nishi; Teppei, Shishido; Keiji, Kimura
• 著者名(英): Xinyi, Li; Masaki, Nishi; Teppei, Shishido; Keiji, Kimura
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: While cloud applications are covering people’s privacy data, their related risks, such as theft risk, leak risk, and other serious risks, are becoming critical issues. Homomorphic encryption (HE) is a promising approach to overcome them since it enables calculation of encrypted data without decryption. Deep learning (DL) inference on cloud is an interesting application of HE. However, the computational cost of HE is too expensive, resulting in its limited usage in the real world. In this paper, we propose an acceleration technique of DL on HE, and implement it in HE-Transformer, which is a framework for DL on HE. Our approach is reducing the bit-width of internal data structure of HE libraries, SEAL and HEXL, from 64-bit to 32-bit. As a result, we can obtain up to 3.59× speedup compared to the original version, while still keeping data privacy.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: While cloud applications are covering people’s privacy data, their related risks, such as theft risk, leak risk, and other serious risks, are becoming critical issues. Homomorphic encryption (HE) is a promising approach to overcome them since it enables calculation of encrypted data without decryption. Deep learning (DL) inference on cloud is an interesting application of HE. However, the computational cost of HE is too expensive, resulting in its limited usage in the real world. In this paper, we propose an acceleration technique of DL on HE, and implement it in HE-Transformer, which is a framework for DL on HE. Our approach is reducing the bit-width of internal data structure of HE libraries, SEAL and HEXL, from 64-bit to 32-bit. As a result, we can obtain up to 3.59× speedup compared to the original version, while still keeping data privacy.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AN10096105
• 書誌情報: 研究報告システム・アーキテクチャ（ARC）; 巻 2022-ARC-249, 号 15, p. 1-6, 発行日 2022-07-20
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8574
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会