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

An Upload-Efficient Scheme for Transferring Knowledge From a Server-Side Pre-trained Generator to Clients in Heterogeneous Federated Learning

CVPR 2024 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Non-IID federated learning · 30 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 FedKTL, an upload-efficient heterogeneous federated learning scheme that uses ETF classifiers on clients, a server-side feature transformer mapping prototypes into a pre-trained StyleGAN-XL latent space, and an extra supervised task on generated prototypical image-vector pairs to transfer generator knowledge to heterogeneous client models.

Problem

Knowledge sharing in heterogeneous federated learning is hard because of data heterogeneity and model heterogeneity; existing methods rely on global datasets, auxiliary models (costly to upload), or class prototypes (biased and carrying limited knowledge).

Approach

Use ETF classifiers to obtain unbiased class prototypes, upload only C×K prototype values, have a server-side trainable feature transformer map prototypes into the valid latent W space of a frozen pre-trained StyleGAN-XL generator, generate one image per class, and let clients train an extra supervised task on these image-vector pairs to absorb the generator's knowledge.

Key results

• FedKTL reaches 88.43% on CIFAR-10 (pathological) and 62.01% on CIFAR-100 (pathological) with HtFE8, outperforming the seven baselines.
• FedKTL outperforms baselines by up to 7.31% in accuracy with HtFE9 model heterogeneity (per the paper's Sec. 4.3).
• FedKTL has the lowest upload cost (0.09M per iteration on CIFAR-100 with 20 clients), since K=C.

Measured effects reported in the paper

Surprising findings

• Using a conditional StyleGAN-XL to directly generate class-wise image-vector pairs without the feature transformer slightly hurts (0.26% drop vs. no collaboration), showing that adapting the generator's latent domain is necessary.
• Baseline methods such as FedKD and FML degrade with more client training epochs, while FedKTL stays stable.

Limitations

• Download cost (7.17M) is much larger than other methods' cost because it transmits generated images.
• Relies on a publicly available pre-trained generator whose domain must be aligned with client tasks; results depend on generator choice and the hyperparameter lambda re-tuned per generator.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 Flowers102 Tiny-ImageNet ImageNet

Benchmarks: CIFAR-10 CIFAR-100 Flowers102 Tiny-ImageNet

Metrics: test accuracy upload cost download cost

Baselines: LG-FedAvg FedGen FedGH FML FedKD FedDistill FedProto

Builds on: FedETF FedProto StyleGAN-XL ArcFace FedGen

Closest papers by meaning

• Towards Diverse Device Heterogeneous Federated Learning via Task Arithmetic Knowledge Integration (NeurIPS 2024 · similarity 0.88)
• Knowledge Transfer via Compact Model in Federated Learning (Student Abstract) (AAAI 2024 · similarity 0.88)
• FedLPS: Heterogeneous Federated Learning for Multiple Tasks with Local Parameter Sharing (AAAI 2024 · similarity 0.87)
• Efficient Federated Learning via Clients-to-Server Knowledge Distillation (Student Abstract) (AAAI 2025 · similarity 0.87)
• FedDAT: An Approach for Foundation Model Finetuning in Multi-Modal Heterogeneous Federated Learning (AAAI 2024 · similarity 0.87)
• FedCompass: Efficient Cross-Silo Federated Learning on Heterogeneous Client Devices Using a Computing Power-Aware Scheduler (ICLR 2024 · similarity 0.87)