Rewarding the Rare: Maverick-Aware Shapley Valuation in Federated Learning

Abstract

Federated Learning (FL) allows clients to train a model collaboratively without sharing their private data. Shapley value (SV) provides a principled way to quantify client contributions in FL. However, existing SV methods use uniform per-class weighting during validation, treating all classes as equally important. This uniform weighting breaks down in the presence of clients with underrepresented or rare classes, also referred to as Mavericks. Such clients are often undervalued due to lower model performance on these challenging classes, despite their critical role in improving generalization. To address this, we introduce a Maverick-aware Shapley valuation framework that reweights validation scores based on per-class accuracy, assigning greater importance to classes where models perform poorly. Building on this, we design FedMS, a Maverick-Shapley client selection mechanism that leverages our refined contribution scores to guide intelligent client selection. Experiments on benchmark datasets demonstrate that FedMS improves model performance and better recognizes valuable client contributions, even under scenarios involving adversaries, free-riders, and skewed or rare-class distributions.