Data-Free Client Contribution Estimation in Federated Learning via Gradient von Neumann Entropy

A novel technique has been developed for estimating client contributions in Federated Learning that does not depend on validation datasets or self-reported data. This method leverages the matrix von Neumann (spectral) entropy of gradient updates from the final layer to assess the diversity of the information provided. Two practical implementations are suggested: SpectralFed, which employs normalized entropy as weights for aggregation, and SpectralFuse, which integrates entropy with class-specific alignment through a rank-adaptive Kalman filter to ensure stability in each round. Tested against benchmarks like CIFAR-10/100, FEMNIST, and FedISIC, the scores derived from entropy exhibit a strong correlation with individual client accuracy across various non-IID data distributions, eliminating the need for validation data or client metadata.