Neural Clipping for Differentially Private SGD

Bayesian optimization for discovering adaptive gradient clipping functions that improve convergence in DP-SGD while maintaining provable (ε, δ)-differential privacy guarantees.

Setup

conda env create -f environment.yml
conda activate neural_clipping

Search Script

python scripts/run_search.py --gpu --config configs/search_config.yaml

This runs 2000 trials of Bayesian optimization using multivariate TPE.