Federated learning experiments with differential privacy.

This directory contains source code for training of six federated optimization tasks with differential privacy including quantile-based adaptive clipping as described in "Differentially Private Learning with Adaptive Clipping" (2021).

This library uses TensorFlow Federated. For a more general look at using TensorFlow Federated for research, see Using TFF for Federated Learning Research.

Some pip packages are required by this library, and may need to be installed:

pip install --requirement differential_privacy/requirements.txt

Example usage

bazel run run_federated -- \
  --client_optimizer=sgd \
  --server_optimizer=sgd \
  --server_sgd_momentum=0.9 \
  --clients_per_round=10 \
  --uniform_weighting=True \
  --clip=0.1 \
  --target_unclipped_quantile=0.5 \
  --adaptive_clip_learning_rate=0.2 \
  --noise_multiplier=0.1 \
  --task=stackoverflow_word \
  --client_learning_rate=0.3 \
  --server_learning_rate=3 \
  --total_rounds=1500 \
  --client_batch_size=16 \
  --root_output_dir=/tmp/dp \
  --experiment_name=so_nwp \
  --max_elements_per_client=1000