Research map / Federated learning and privacy / Personalized and heterogeneous FL

Cross-Silo Feature Space Alignment for Federated Learning on Clients with Imbalanced Data

AAAI 2025 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Non-IID federated learning · 15 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FedFSA, a federated learning method that uses CLIP text embeddings to regularize local training, transfers variance from majority to minority classes to build prototypical spaces, and aligns cross-client features on the server to improve aggregation under data imbalance.

Problem

Data imbalance across clients in federated learning yields inconsistent local feature spaces, causing ill-posed model aggregation and hurting the global model's generalization.

Approach

Combines in-silo CLIP-text-guided prototypical contrastive learning with variance-transfer augmentation of minority-class prototypes, and then on the server aligns features from different clients into a shared space via local/complementary consistency losses and importance-weighted classification, as a model-agnostic plug-in to existing FL methods.

Key results

• FedFSA on FedAvg reaches 74.45% on CIFAR10 with K=5 clients, versus 70.85% for FedAvg and 73.24% for FedCSPC.
• FedFSA with FedETF reaches 75.15% on CIFAR10 (K=5) and 64.23% on CIFAR100 (K=5) and 49.16% on TinyImageNet (K=10).
• Ablation on CIFAR10 (K=5) shows the full model at 74.45% versus 72.06% with TERL only and 73.21% with TERL+CSFSA.

Measured effects reported in the paper

Surprising findings

• Model calibration is suggested to help personalized local models by leveraging other clients' knowledge, though the paper reports this qualitatively.

Limitations

• Calibration may still fail when local learning is unreliable, as shown in the error analysis.
• Evaluation limited to three image classification datasets with small CNN/ResNet18 backbones and Dirichlet beta=0.5.
• Excessive variance fusion can cause inter-class overlap and degrade performance.

Experimental setup

Datasets: CIFAR10 CIFAR100 TinyImageNet

Benchmarks: CIFAR10 CIFAR100 TinyImageNet

Metrics: Top-1 accuracy

Baselines: FedAvg MOON FedProc FedDecorr FedETF FedRCL CCVR FedCSPC CLIP2FL

Builds on: MOON FedProc CCVR FedETF CLIP2FL

Closest papers by meaning

• FedLoGe: Joint Local and Generic Federated Learning under Long-tailed Data (ICLR 2024 · similarity 0.84)
• Taming Cross-Domain Representation Variance in Federated Prototype Learning with Heterogeneous Data Domains (NeurIPS 2024 · similarity 0.84)
• Classifier Clustering and Feature Alignment for Federated Learning under Distributed Concept Drift (NeurIPS 2024 · similarity 0.84)
• PICSR: Prototype-Informed Cross-Silo Router for Federated Learning (Student Abstract) (AAAI 2024 · similarity 0.84)
• Tackling Feature-Classifier Mismatch in Federated Learning via Prompt-Driven Feature Transformation (NeurIPS 2025 · similarity 0.83)
• FedSA: A Unified Representation Learning via Semantic Anchors for Prototype-based Federated Learning (AAAI 2025 · similarity 0.83)