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

A Simple Data Augmentation for Feature Distribution Skewed Federated Learning

CVPR 2025 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Non-IID federated learning · 13 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 FedRDN, a non-parametric input-level augmentation that randomly normalizes each image with the channel-wise statistics of clients across the federation, improving multiple FL methods under feature distribution skew.

Problem

Feature distribution skewed federated learning, where clients' data have different underlying distributions causing feature shift and degraded global performance.

Approach

Extends standard data normalization to FL by randomly sampling the mean and std of some client's local dataset (shared as aggregated channel-wise statistics) to normalize each training image, injecting federation-wide distribution information at the input level without changing the network or adding much communication; at test time, each client uses its own statistics.

Key results

• FedRDN improves FedAvg on Office-Caltech-10 from 62.51% to 69.80% average accuracy (+7.29), and gains up to 11.21% for Scaffold.
• On ProstateMRI, FedRDN raises FedFA's average Dice from 91.90 to 93.14 (+1.24).
• FedRDN-V (fixed average statistics, i.e., conventional normalization) performs notably worse than the random-statistics variant, e.g., 61.46 vs 69.80 on Office-Caltech-10 for FedAvg.

Measured effects reported in the paper

Surprising findings

• Conventional data normalization with averaged statistics often decreases FL performance, while random selection of client statistics consistently helps.

Limitations

• Focuses on visual tasks; authors argue data statistics are privacy-agnostic but no formal privacy analysis is given.
• Mainly empirical; lacks theoretical analysis of why the method works.
• Gains on some methods (e.g., FedBN on ProstateMRI, +0.49 on Office-Caltech-10) are modest.

Experimental setup

Datasets: Office-Caltech-10 DomainNet ProstateMRI

Benchmarks: Office-Caltech-10 DomainNet ProstateMRI

Metrics: top-1 accuracy Dice coefficient

Baselines: FedAvg FedAvgM FedProx Scaffold FedNova FedBN FedProto FedFA FedMix conventional data normalization

Builds on: FedAvg FedMix FedFA FedBN

Closest papers by meaning

• FedWon: Triumphing Multi-domain Federated Learning Without Normalization (ICLR 2024 · similarity 0.89)
• Tackling Feature-Classifier Mismatch in Federated Learning via Prompt-Driven Feature Transformation (NeurIPS 2025 · similarity 0.88)
• Federated Learning with Domain Shift Eraser (CVPR 2025 · similarity 0.88)
• FLUX: Efficient Descriptor-Driven Clustered Federated Learning under Arbitrary Distribution Shifts (NeurIPS 2025 · similarity 0.87)
• Federated Causally Invariant Feature Learning (AAAI 2025 · similarity 0.87)
• FedAvP: Augment Local Data via Shared Policy in Federated Learning (NeurIPS 2024 · similarity 0.87)