search

aidecentralized / sonar

SONAR - Self-Organizing Network of Aggregated Representations
MIT License
13 stars 34 forks source link

Training with DPSGD Implementation #113

Closed hardlyhuman closed 1 week ago

hardlyhuman commented 1 month ago

This pull request introduces support for training models using the DPSGD (Differentially Private Stochastic Gradient Descent) algorithm, ensuring differential privacy during training. DPSGD allows the model to learn while preserving the privacy of individual training examples, making it suitable for sensitive data scenarios.

tremblerz commented 1 month ago
  1. This looks good but I guess you should use a library (or write on your own) for privacy accounting or ideally pass epsilon as a parameter in the config.
  2. Also, can you update the config and put "if/else" wherever necessary so that anyone can use the dpsgd routine.
tremblerz commented 1 week ago

Inactive by the author so closing it. We will probably do it sometime soon using Opacus