Title: Bridging the Alignment-Maturity Gap in Federated Prototype Learning

Abstract

Deriving discriminative visual features from distributed, non-homogeneous datasets remains a core obstacle in Federated Learning (FL). While prototype-based approaches mitigate statistical heterogeneity by exchanging class-level representations among clients, they inadvertently induce a distance-dependent gradient pressure. This phenomenon is most acute during the initial training phases: the alignment force exerted on nascent global prototypes—compiled from noisy local embeddings—produces excessive gradients that stifle the development of local discriminative patterns. Consequently, the resulting embedding space lacks organization, leading to diminished recognition accuracy, a problem that intensifies under severe non-IID conditions.

To address these issues, we introduce FedSAP, a framework designed to stabilize federated representation learning via two synergistic mechanisms. First, it employs a deterministic alignment curriculum that postpones global alignment until local representations have achieved stability. Second, it utilizes a geometry-driven proxy separation loss to enforce inter-class structure on the unit hypersphere by leveraging the existing prototype bank, thereby avoiding any increase in parameters or communication costs. These strategies yield compact, distinctly separated class clusters while leaving the fundamental communication protocol between federation participants unchanged.

Our evaluation across three benchmarks, covering a spectrum of heterogeneity levels, demonstrates performance improvements of up to 4 percentage points compared to existing prototype-based baselines, with the most significant gains observed in highly heterogeneous environments. Furthermore, the representational foundation of our framework allows for a seamless adaptation to semi-supervised scenarios, where unlabelled data is integrated with minimal structural changes. This versatility highlights the broad applicability of scheduled alignment as a core design principle.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC