Beyond Leave-One-Out: Private and robust contribution evaluation in Federated Learning

Artikel i vetenskaplig tidskrift, 2026

Cross-silo Federated Learning enables multiple organizations to collaboratively train machine learning models without sharing raw data, but client updates may still leak sensitive information through inference attacks. Secure Aggregation protects privacy by concealing individual updates; however, it also complicates contribution evaluation, which is essential for fair reward allocation and identifying low-quality or malicious participants. Existing marginal-contribution approaches, such as the Shapley value, are incompatible with Secure Aggregation, while practical alternatives, such as Leave-One-Out, are limited and rely on self-evaluation. We introduce two marginal difference based contribution scores compatible with Secure Aggregation. Fair-Private satisfies standard fairness axioms, while Everybody-Else eliminates self-evaluation and provides resistance to direct manipulation, addressing an overlooked vulnerability in contribution evaluation. We establish theoretical guarantees for fairness, privacy, robustness, and computational efficiency, and evaluate our methods on multiple medical image datasets and CIFAR10 in cross-silo settings. Our scores consistently outperform existing baselines, more closely approximate Shapley-based client rankings, and improve downstream model performance and misbehavior detection. On ISIC2019, for example, they reduce the ranking error relative to the Shapley value by an order of magnitude compared to Leave-One-Out (0.11 vs. 1.06) and improve the Spearman correlation from 0.82 to 0.94. These results demonstrate that fairness, privacy, robustness, and practical utility can be jointly achieved in federated contribution evaluation, providing a principled solution for real-world cross-silo deployments.