Data Valuation and Detections in Federated Learning

Federated Learning (FL) enables collaborative model training without sharing raw data, demanding abundant, high-quality data for optimal model performance. Fair and efficient data evaluation is a fundamental issue for incentivizing clients to provide more high-quality data. Meanwhile, it is likely that only a subset of clients and datasets are relevant for a learning task while the rest of them may have a negative impact on the model training. This paper introduces a novel privacy-preserving method for evaluating client contributions and selecting relevant data samples without a pre-specified training algorithm. Our proposed approach, FedBary, utilizes Wasserstein distance within the federated context, offering a new pioneering solution for data valuation, which provides transparent data evaluation and efficient computation of Wasserstein barycenter to mitigate reliance on validation data. We conduct extensive empirical experiments and theoretical analysis, showing the promising research of this valuation metric.