StabFL: Stabilized Logit Adjustment with Queue-Augmented Representation Learning for Label-Skewed Federated Learning

Federated learning under label skew aggregates models trained with different local class distributions. Local logit adjustment can reduce class-dependent bias, but the adjusted loss updates both the encoder and classifier in client-specific directions. StabFL optimizes a joint local objective that combines logit-adjusted classification loss, full-model proximal regularization, and supervised contrastive learning with a client-local queue. The queue retains recent labeled projected representations and extends the comparison set available to small mini-batches. Across CIFAR-10 and PAMAP2, StabFL reaches 85.57±1.36% and 75.76±3.21% accuracy on CIFAR-10 under moderate and severe label skew, and 83.21±0.46% and 74.68±4.10% on PAMAP2 under the same skew levels. Compared with StabFL w/o Queue, the client-local queue raises mean accuracy by 0.39 and 1.66 percentage points on CIFAR-10 and by 1.76 and 4.44 percentage points on PAMAP2 while transmitting no queued projected representations, labels, or local class priors.