Fetching metadata from the HF Docker repository...
-
Fix DP-SGD implementation and add real-time training progress
Hutches67 +1: Cursor Major changes:
- Implement correct DP-SGD noise formula based on research (Optax/TF Privacy)
- Add Server-Sent Events (SSE) for real-time epoch-by-epoch training progress
- Remove parameter capping to respect user-specified privacy settings
- Update presets with research-validated parameters (~95-97% MNIST accuracy)
- Add rate limiting for training endpoint
- Consolidate gradient utilities into shared module
- Improve privacy calculator with RDP-based accounting
- Fix security headers and CORS configuration
- Add threaded mode for Flask SSE streaming support
Research-validated defaults:
- noise_multiplier=1.1, clipping_norm=1.0, learning_rate=0.15
- Achieves ~96% accuracy on MNIST with reasonable privacy (ε≈3-5)
Co-authored-by: Cursor <cursoragent@cursor.com>