DIV-1 · Divanshu Sharma

Work / Fig. 02 · ML research · Yale University · 2025–26

Privacy-Preserving Deep Learning via MPC

88.08% test accuracy — ChestMNIST, strided-convolution CNN, 5 epochs @ batch 512.

Fig. 02

Privacy-Preserving Deep Learning via MPC

ML research · Yale University · 2025–26

Problem

Institutions holding sensitive data — hospitals, biobanks — cannot pool it to train shared models. Standard frameworks assume the training data is visible to the machine doing the training.

Approach

Engineered a secure, open-source deep learning library from first principles in Python/NumPy — no external autograd — around a compiler-centric multi-party computation architecture, so collaborative training carries verifiable cryptographic guarantees rather than policy promises. Proposed the CNN stack upstream to Sequre (0xTCG/sequre) through a series of pull requests.

Results

  • 88.08% test accuracy

    ChestMNIST, strided-convolution CNN, 5 epochs @ batch 512

  • 0 external autograd deps

    backprop, conv & pooling layers derived and implemented by hand

  • 4 upstream PRs

    CNN layers & MPC training pipeline proposed to 0xTCG/sequre

Release v2.0 · Oct 2025 — Mar 2026

ML Researcher — Yale University

  • 88.08% ChestMNIST test accuracy in 5 epochs at batch size 512 — trained under MPC.
  • No external autograd: layers, backprop and optimizers derived by hand.
  • CNN stack proposed upstream to 0xTCG/sequre across four pull requests.