18 October 2026 in SSRN Electronic Journal
Label skew poses a major challenge to federated learning by inducing dimensional collapse and inconsistent decision geometry across heterogeneous clients. Existing methods largely address these problems in isolation: feature decorrelation mitigates dimensional collapse, whereas classifier calibration or feature alignment promotes a more consistent decision geometry. However, these two problems are intrinsically coupled. Expanding a collapsed representation space without preserving class structure does not necessarily improve discrimination, while imposing a shared classifier geometry on a spectrally deficient representation limits the capacity available for class separation. To jointly address these issues, we propose FedBlade, a unified framework that couples accelerated feature decorrelation with prototype-guided bilateral alignment. Specifically, LDDecorr maximizes the log-determinant of the representation correlation matrix, yielding inverse-spectrum gradients that exert stronger restoring pressure on small singular values and thereby accelerate the mitigation of dimensional collapse. Meanwhile, PBA uses global prototypes as bridges among the feature extractor, projector, and shared ETF classifier, jointly aligning the representation and projection spaces toward a globally consistent class geometry. These two mechanisms are complementary: LDDecorr preserves the representational capacity required for effective geometric alignment, while PBA constrains the recovered capacity toward discriminative class structure. Extensive experiments demonstrate that FedBlade consistently achieves superior performance across diverse label-skew settings while substantially accelerating model convergence.
1 September 2026 in IEEE Transactions on Services Computing