FedNPC: Stochastic Noise-Driven Post-hoc Classifier Calibration Method for Federated Long-Tailed Learning

Jintong Gao, He Zhao, Yibo Yang, Dandan GuoCVPR 2026 Findings 2026 · OpenReview

TL;DR

FedNPC introduces a data-free, post-hoc calibration method for Federated Long-Tailed Learning (Fed-LT) that fixes classifier bias using nothing but pure isotropic Gaussian noise. By leveraging the gradient dynamics of noise on a frozen feature extractor, it equalizes classifier weight norms, significantly boosting accuracy on tail classes while preserving client privacy and minimizing communication overhead.

What Problem Does This Paper Address?

Real-world Federated Learning (FL) faces a double whammy: local data heterogeneity (clients have wildly different data distributions) and a global long-tailed distribution (a few "head" classes have tons of samples, while most "tail" classes have very few). This combination is known as Federated Long-Tailed Learning (Fed-LT).

When models are trained under these conditions, the classifier becomes heavily biased. The authors discover that the root cause is a severe imbalance in the classifier's weight norms ($||w_c||_2$)—head classes end up with massive norms, completely drowning out the tail classes. Crucially, this norm imbalance is significantly worse in federated settings than in centralized ones.

Figure 1. Empirical experiments on CIFAR-100 test dataset with ResNet-8 as the backbone, where imbalance factor is IF=100 under different non-IIDness α. (a) Accuracy and (b) classifier weight norms ||wc||2 trained under federated long-tailed (Fed-LT, FedAvg), centralized long-tailed (LT), and centralized balanced settings across class groups (averaged per 10 classes).

Existing calibration methods try to fix this on the server, but they usually require an auxiliary balanced dataset (violating the data-free server principle) or force clients to upload feature statistics or gradients, which introduces privacy risks and heavy communication overhead.

Methods

The authors propose FedNPC (Stochastic Noise-driven Post-hoc Classifier Calibration), a lightweight, data-free approach that operates entirely server-side in two stages:

  1. Stage 1: Standard Federated Training: Train the global model using any standard FL or Fed-LT algorithm. This yields a high-quality feature extractor but a highly biased classifier.
  2. Stage 2: Noise-driven Post-hoc Calibration (NPC): On the server, freeze the feature extractor. Construct a virtual, perfectly balanced dataset where inputs are sampled from isotropic Gaussian noise ($z \sim \mathcal{N}(0, I_d)$) and labels are drawn from a uniform distribution ($y \sim \mathcal{U}(C)$). Fine-tune only the classifier weights ($W_g$) by minimizing cross-entropy loss on this virtual dataset.
The two-stage training process of FedNPC.

Why does noise work?

The authors prove via expected gradient dynamics that isotropic noise acts as an adaptive norm equalizer. Head classes (large weight norms) generate high logit variance, which gives them an advantage in the softmax competition. When exposed to uniform labels, the optimization process naturally applies negative feedback to these large norms (shrinking them) and positive feedback to small norms (growing them), driving the classifier toward a perfectly balanced state.

Datasets

The method was evaluated on three major long-tailed benchmarks:

  • CIFAR-10-LT & CIFAR-100-LT: Artificially imbalanced versions of CIFAR using an exponential distribution with Imbalance Factors (IF) of 10, 50, and 100. Local client data was split using a Dirichlet distribution ($\alpha = 0.1, 0.5, 1.0$) to simulate non-IIDness.
  • ImageNet-LT: A practical long-tailed subset of ImageNet containing 1,000 classes with an imbalance factor of IF=256, tested under severe heterogeneity ($\alpha = 0.1$).

Key Insights and Ideas

  • The Heterogeneity Amplification Effect: Local data heterogeneity doesn't just add noise; it actively exacerbates global long-tail class bias, leading to a much steeper drop-off in tail-class accuracy compared to centralized long-tailed learning.
  • Isotropy Over Semantics: You don't need real, semantically meaningful data to calibrate a classifier. A statistically pure, isotropic proxy (Gaussian noise) is actually better because real data carries anisotropic bias ($\Sigma \neq \sigma^2 I$) that can interfere with perfect norm equalization.
  • Mathematical Self-Correction: The softmax function naturally penalizes high variance when paired with uniform labels. Because large weight norms cause high logit variance, the gradient dynamics automatically self-correct the classifier's structural imbalance.

Key Results

  • Substantial Accuracy Boosts: On CIFAR-10-LT and CIFAR-100-LT ($\alpha=0.5$), FedNPC provided an average accuracy improvement of 5.37% when applied to standard FL frameworks (like FedAvg, FedProx, FedLWS) and 3.89% when paired with dedicated Fed-LT methods.
  • Massive Gains on Head-Heavy Setups: When applied to standard FedAvg on CIFAR-10-LT (IF=100), FedNPC boosted accuracy from 56.36% to 68.52%—a massive 12.16% absolute jump.
  • Large-Scale Generalization: On the ImageNet-LT benchmark with a ResNet-50 backbone, FedNPC consistently improved performance across all baselines, notably boosting FedLWS overall accuracy by 1.46% and significantly improving few-shot (tail) classes.
  • Zero Overhead: It achieves these gains completely server-side with zero extra client communication, data sharing, or statistical uploads.

Limitations and Future Work

  • Diminishing Returns on Simple Datasets: While FedNPC works exceptionally well on complex datasets like CIFAR-100-LT, its performance gains are more marginal on simpler long-tailed datasets with low imbalance factors.
  • Dependence on Stage 1 Features: The method assumes that Stage 1 yields a sufficiently high-quality feature extractor. If the feature representation itself is severely compromised by extreme non-IID settings, classifier-only calibration might hit a performance ceiling.
  • Future Directions: The authors suggest extending this noise-driven calibration paradigm to other imbalanced federated tasks, such as federated object detection or semantic segmentation.

Summary updated

Abstract

Federated long-tailed learning (Fed-LT) simultaneously tackles data heterogeneity across clients and global long-tailed distribution challenges. Existing approaches often depend on auxiliary datasets, gradients, or statistical information, which not only increase communication overhead but also introduce significant privacy leakage risks. To address these issues, we first conduct a systematic empirical study of the classifier bias under imbalanced data. We observe that the norm imbalance problem is substantially more severe in Fed-LT than in the centralized LT setting. Motivated by this observation, we propose FedNPC (stochastic Noise-driven Post-hoc classifier Calibration), a novel data-free post-processing calibration method for Fed-LT, using a virtual, balanced dataset. Furthermore, we provide theoretical insights showing that isotropic noise can effectively mitigate classifier norm imbalance. Extensive experiments on multiple Fed-LT benchmarks demonstrate the superior performance and generalization ability of FedNPC. Our code is available athttps://github.com/JintongGao/FedNPC.
BibTeX
@inproceedings{gao2026fednpc,
  author = {Jintong Gao and He Zhao and Yibo Yang and Dandan Guo},
  title = {{FedNPC: Stochastic Noise-Driven Post-hoc Classifier Calibration Method for Federated Long-Tailed Learning}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2026},
  url = {https://openreview.net/forum?id=4befwY9BMr},
}
Figure