DP-Hype: Federated Differentially Private Hyperparameter Search
Researchers have developed DP-Hype, a new algorithm for federated hyperparameter search that incorporates differential privacy. This method allows clients in a federated learning setup to collectively select hyperparameters through a voting mechanism based on local evaluations, ensuring a compromise supported by a majority. DP-Hype guarantees client-level differential privacy without being dependent on the number of hyperparameters and offers utility bounds, demonstrating its effectiveness even with small privacy budgets. AI
IMPACT Enhances privacy in federated learning by enabling secure hyperparameter tuning.