Data-Free Client Contribution Estimation via Logit Maximization for Federated Learning

2026-05-17

AI Summary

This work addresses the challenge in federated learning where class imbalance and label skew often cause model aggregation to be dominated by clients with abundant labels, thereby degrading performance on minority classes. To mitigate this without requiring raw data, metadata, or auxiliary datasets, the authors propose CELM—a class-aware client contribution evaluation and aggregation framework. CELM leverages server-side logit maximization probing of client updates to construct a cross-client evidence matrix, then computes class-level contribution weights through simplex constraints and momentum smoothing to enhance discriminability for underrepresented classes. Experiments demonstrate that CELM significantly improves model robustness against class imbalance and statistical heterogeneity across diverse non-IID and extreme label-partitioned visual benchmarks, all without incurring additional communication overhead.

Abstract

Federated learning (FL) enables collaborative learning of computer vision models, where privacy and regulatory constraints prevent centralizing data across devices or organizations. However, practical FL deployments often exhibit severe class imbalance and label skew, causing standard aggregation protocols to overfit dominant clients and degrade minority-class performance. We propose a data-free, class-wise contribution estimation and aggregation framework based on logit maximization (CELM) that does not require sharing raw data, client metadata, or auxiliary public datasets. The FL server probes client updates to obtain class-wise evidence scores and assembles a cross-client evidence matrix, which quantifies both per-class competence and class coverage. Using this matrix, we compute contribution weights that upweight clients providing strong, discriminative evidence for underrepresented classes. The resulting aggregation is stable due to simplex constraints and momentum smoothing, and it remains compatible with standard FL training pipelines. We evaluate the approach on representative vision benchmarks under controlled non-IID and pathological label splits, demonstrating that CELM-based aggregation improves robustness to imbalance and statistical heterogeneity, while yielding better performance without requiring any additional data exchange.

Problem

• federated learning
• class imbalance
• label skew
• client contribution
• non-IID

Innovation

• data-free
• logit maximization
• client contribution estimation
• federated learning
• class imbalance