AHEC: End-to-end Compiler Framework for Privacy-preserving Machine Learning Acceleration

@inproceedings{chen2020ahec,
  title={Ahec: End-to-end compiler framework for privacy-preserving machine learning acceleration},
  author={Chen, Huili and Cammarota, Rosario and Valencia, Felipe and Regazzoni, Francesco and Koushanfar, Farinaz},
  booktitle={2020 57th ACM/IEEE Design Automation Conference (DAC)},
  pages={1--6},
  year={2020},
  organization={IEEE}
}

• first end-to-end PPML HE Compiler framework

2. Preliminaries

• HE
  • 加算と乗算で準同型の話
  • leveled FHEに焦点
• HE Compiler for PPML
  • TVM
  • nGraph-HE
• HE-based PPML
  • CryptoNets
  • nGraph-HE
  • CHET
• HE + Garbled Circuit PPML
  • Gazelle

3. AHEC Framework

• PlaidML を拡張
• CKKS scheme of Microsoft SEAL を最適化

4. Evaluation

2022-12-14

HE operation-level

• Microsoft SEAL の op が早くなった
  • どうやって? todo

ML kernel-level

• Dot Product
  • N = 32, 64, 128, 256
• GEMM: General Matrix Multiplication
  • 32x32 - 32x16, 64x64 - 64x16, 128x128 - 128-16, 256x256 - 256x16
• Conv: Convolution
  • filter: 3x3
  • image: 28x28, 32x32, 64x64, 256x256

• neural network layers