How to Privately Tune Hyperparameters in Federated Learning? Insights from a Benchmark Study

TL;DR

The paper tackles privacy-preserving hyperparameter tuning in cross-silo federated learning by first benchmarking how locally discovered HPs can be aggregated at the server under IID and non-IID data. It finds that simple averaging suffices for IID settings, while density-based clustering robustly recovers optimal server HPs in non-IID scenarios. To safeguard client HPs from leakage, it then introduces PrivTuna, a multiparty homomorphic encryption-based framework implementing privacy-preserving aggregation strategies, notably PF-Mean and PF-DBSCAN. Empirical results demonstrate that PrivTuna achieves efficient computation and communication with minimal distortion in server HP tuning, enabling private, accurate configuration of FFPL hyperparameters prior to training. Overall, the work provides a practical pathway for private, single-shot server HP tuning in PPFL via server-client HP aggregation and encrypted computation.

Abstract

In this paper, we address the problem of privacy-preserving hyperparameter (HP) tuning for cross-silo federated learning (FL). We first perform a comprehensive measurement study that benchmarks various HP strategies suitable for FL. Our benchmarks show that the optimal parameters of the FL server, e.g., the learning rate, can be accurately and efficiently tuned based on the HPs found by each client on its local data. We demonstrate that HP averaging is suitable for iid settings, while density-based clustering can uncover the optimal set of parameters in non-iid ones. Then, to prevent information leakage from the exchange of the clients' local HPs, we design and implement PrivTuna, a novel framework for privacy-preserving HP tuning using multiparty homomorphic encryption. We use PrivTuna to implement privacy-preserving federated averaging and density-based clustering, and we experimentally evaluate its performance demonstrating its computation/communication efficiency and its precision in tuning hyperparameters.

Figures (8)

• Figure 1: Our HP tuning method for cross-silo federated learning. ${hp}_k$ and ${hp}_g$ denote the local best HP configuration at client $k$ and the global best configuration (denoted by $g$), respectively.
• Figure 2: Barplots of learning rate, momentum, and accuracy for the iid setting with Mean as the $\textsf{Combine}(\cdot)$ strategy, on various datasets and experiments. The ground truth (GHO result) is indicated by black dots. The remaining bar colors represent the results of combining the client optimal HPs with the Mean$\textsf{Combine}(\cdot)$ strategy, for variable number of clients ($N$).
• Figure 3: Barplots of learning rate and momentum for the non-iid setting with $N=20$ clients, variable datasets and $\textsf{Combine}(\cdot)$ strategies. The top-row results are for feature skew ($\beta_f=0.02$), middle-row results for label skew ($\beta_{\ell}=1.0$), and bottom-row results are for quantity skew ($\beta_q=0.4$). Bar colors represent the HPs derived using various combination strategies.
• Figure 4: Barplots of learning rate, momentum, and accuracy, for the non-iid setting with label skew ($\beta_{\ell}=1.0$), using DBSCAN as the $\textsf{Combine}(\cdot)$ strategy on various datasets and experiments. The ground truth (GHO results) is indicated by black dots. Bar colors represent the results of combining the client optimal HPs with the DBSCAN strategy, for variable number of clients ($N$).
• Figure 5: Comparison of the various combination strategies proposed in FLoRA flora vs. the density-based clustering strategy identified in our measurement study. The blue bar indicates the federated grid search results as ground truth. Each plot represents a different skew type (averaged across skew parameters with 10 and 20 clients).