CrypTorch: New Compiler Optimizes MPC-Based Machine Learning
Researchers have introduced CrypTorch, a compiler framework designed to improve the performance and accuracy of machine learning models that use multi-party computation (MPC). MPC allows multiple parties to run ML workloads without sharing private data or model parameters, but existing frameworks often degrade accuracy and performance due to opaque MPC-specific transformations. CrypTorch addresses this by splitting transformations into modular, inspectable stages and emitting executable graphs after each step for iterative testing. It automatically selects optimal transformations from a pool of choices to balance trade-offs. The framework is modular, extensible, and iteratively testable, aiming to make MPC-based ML more efficient and transparent. The work is detailed in arXiv paper 2511.19711.
Key facts
- CrypTorch is a compiler framework for MPC-based machine learning.
- MPC-based ML runs workloads across multiple parties without sharing private data.
- Existing frameworks often degrade accuracy and performance due to opaque transformations.
- CrypTorch splits transformations into modular compilation stages.
- Users can inspect and optimize transformations with CrypTorch.
- CrypTorch emits an executable graph after each transformation for iterative testing.
- It automatically selects transformations from a pool to balance trade-offs.
- The paper is available on arXiv with identifier 2511.19711.
Entities
Institutions
- arXiv