Calculating epsilon budget for the federated model with DPquery

[Henrique-Potter]

Hi is it possible to get the epsilon budget as an evaluation metric in the example from federated/differential_privacy/stackoverflow/run_federated.py? Or how can you calculate it given the parameters for this fed model using the rdp accountant from tensorflow_privacy.privacy.analysis.rdp_accountant?

[galenmandrew]

Yes, the best way is to use the RDP accountant from tensorflow_privacy. Use whatever noise multiplier you ran with and a sampling fraction q=clients_per_round / n. The epsilon computation is not changed whether you are running DP-FedAvg for user-level DP or DP-SGD for example-level DP. Adaptive clipping also does not change the epsilon.