A Universal Metric of Dataset Similarity for Cross-silo Federated Learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Privacy

Background:

• Federated Learning (FL) facilitates collaborative training without raw data sharing, crucial for privacy-sensitive sectors like healthcare.
• Non-Independent and Identically Distributed (non-IID) datasets significantly degrade FL performance.
• Existing dataset similarity metrics for FL collaboration lack interpretability, require direct data access, or are inefficient.

Purpose of the Study:

• To develop a novel, privacy-preserving metric for assessing client dataset similarity in Federated Learning.
• To predict the impact of client collaboration on model performance in FL settings.
• To provide an actionable tool for selecting optimal participants in cross-silo FL.

Main Methods:

• Proposed a novel metric using model representations from a single federated training round.
• Formulated similarity assessment as an optimal transport problem with a hybrid cost function.
• Ensured privacy using Secure Multiparty Computation (SMC) and Differential Privacy (DP).

Main Results:

• The metric accurately predicts collaboration benefits by analyzing feature-level and label distribution divergence.
• Theoretical analysis links the metric to weight divergence, explaining its predictive power.
• Empirical validation shows strong correlation with weight divergence throughout training.

Conclusions:

• The proposed metric reliably identifies beneficial collaborations in Federated Learning.
• This privacy-preserving approach offers an actionable tool for participant selection in cross-silo FL.
• Early-round model representations effectively predict long-term collaboration outcomes.