Federated Learning with Differential Privacy refers to a learning paradigm in which a global model \(f_\theta\) is created by aggregating locally trained models on distributed data sources, while Differential Privacy (DP) ensures that the participation of individual data points is not traceable.
Formally, at each communication step, a DP mechanism \( \mathcal{M} \) is applied to the local updates \( \Delta \theta_i \), e.g., by adding Gaussian noise \( \mathcal{N}(0, \sigma^2) \), such that:
\[ \mathcal{M}(\Delta \theta_i) = \Delta \theta_i + \mathcal{N}(0, \sigma^2 I) \]
The overall privacy is measured via the composition of the individual steps with parameters \( (\varepsilon, \delta) \).
By combining Federated Averaging (FedAvg) with DP mechanisms, the model can be trained without extracting sensitive information from the local data.
This addresses the challenge of mathematically guaranteeing privacy in distributed learning while maintaining high model quality.
Definition:
“Federated Learning with Differential Privacy is a distributed learning procedure in which local model updates are perturbed with mathematically defined noise to protect the privacy of participants while simultaneously training a global model.”
Source:
McMahan, H. B., Ramage, D., Talwar, K., & Zhang, L. (2018). Learning Differentially Private Recurrent Language Models. Proceedings of the 6th International Conference on Learning Representations (ICLR).