How to use the framework with local Differential Privacy

[fraboeni]

How can I use the LDP implementation of the framework?

Thank you very much for the great library. I was trying to use the LDP implementation inside my code. However, I cannot figure out how to really integrate the DP parts into the workflow.

Would it be possible to provide a minimal example that I can use as a starting point to understand how I can pass the parameters to the clients, how I can change the clients to use the DP Optimizer etc.

Thank you very much for your help!

[JohnlNguyen]

Hi you can do do local DP by using DPClient and pass in the base_dp_client to the config

{
  "config": {
    "trainer": {
      "_base_": "base_sync_trainer",
      "epochs": 1,
      "server": {
        "_base_": "base_sync_server",
        "server_optimizer": {
          "_base_": "base_fed_avg_with_lr",
          "lr": 2.13,
          "momentum": 0.9
        },
        "active_user_selector": {
          "_base_": "base_uniformly_random_active_user_selector"
        }
      },
      "client": {
        "_base_": "base_dp_client",
        "epochs": 1,
        "optimizer": {
          "_base_": "base_optimizer_sgd",
          "lr": 0.0138,
          "momentum": 0
        },
        "privacy_setting": {
             "noise_multiplier": 1,
             "clipping": {
                "clipping_value": 1
             }
         }
      },
      "users_per_round": 10,
      "train_metrics_reported_per_epoch": 1,
      "always_keep_trained_model": false,
      "report_train_metrics": true,
      "eval_epoch_frequency": 1,
      "do_eval": true,
      "report_train_metrics_after_aggregation": true
    },
    "data": {
      "local_batch_size": 32,
      "examples_per_user": 500
    }
  }
}

[fraboeni]

Thank you very much @JohnlNguyen!