Uncertainty Quantification in Federated Learning

60 concepts41 papers2 proposals

Standard Federated Learning (FL) algorithms, such as FedAvg, treat decentralized training as a distributed optimization problem. The goal is simple: find a single vector of model weights that minimizes empirical loss across all participating clients.

However, this reliance on a single "best guess"—a point estimate—introduces a severe vulnerability. When local client data is highly heterogeneous (non-IID) or limited in size, standard optimization techniques produce overconfident, miscalibrated models. These models do not know what they do not know. In safety-critical applications like clinical healthcare (Li et al.), autonomous driving, or financial risk assessment, an overconfident incorrect prediction can be catastrophic.

To solve this, a rapidly growing research direction reframes federated learning from distributed optimization to distributed Bayesian inference and uncertainty quantification. Instead of forcing clients to agree on a single set of weights, these frameworks allow clients to collaboratively map out a probability distribution over parameters or predictions. This ensures that when a model encounters out-of-distribution data or an ambiguous input, it signals its uncertainty rather than making a high-confidence error.


Decentralized Posterior Sampling via Langevin Dynamics

How do you map out a complex probability distribution across a network of devices without centralizing their data? One prominent approach adapts Markov Chain Monte Carlo (MCMC) sampling for distributed architectures.

Instead of navigating directly down a local loss surface to a minimum, clients inject a controlled amount of Gaussian noise into their local stochastic gradient updates. This technique, known as Stochastic Gradient Langevin Dynamics (SGLD), forces the model parameters to continuously explore the surrounding probability landscape.

Standard Optimization (Point Estimate):
[Initial Weights] ---> [Gradient Descent] ---> [Single Optimal Weight Vector]

Langevin Dynamics (Posterior Sampling):
[Initial Weights] ---> [Gradient + Gaussian Noise] ---> [Continuous Trajectory / Posterior Samples]

As demonstrated by Deng et al. with Federated Averaging Langevin Dynamics (FA-LD), periodically averaging these noisy local trajectories allows the global system to sample from the true joint posterior distribution. This concept can be extended to fully peer-to-peer networks without any central coordinator, as shown by Gurbuzbalaban et al..

However, data heterogeneity introduces a major challenge: client drift. Because each client's local data implies a different posterior landscape, local sampling chains naturally drift toward local optima. To resolve this, Plassier introduced control variates into the sampling process (VR-FALD*), mathematically eliminating client drift and ensuring that decentralized sampling chains converge accurately to the true global target distribution.


Federated Variational Inference and Particle Methods

While MCMC sampling provides strong theoretical guarantees, transmitting multiple model states (samples) over a network can become communication-heavy. Federated Variational Inference (FVI) bypasses this by turning the sampling problem back into an optimization problem. Instead of saving individual samples, clients optimize the parameters of a known distribution—such as the mean and variance of a Gaussian—to approximate the true posterior.

The authors of early frameworks like DP-PVI (Sharma et al.) proved that this probabilistic framing naturally aligns with other decentralized constraints. For instance, the noise injection required for differential privacy can be integrated directly into variational updates without destroying the model's capacity to output principled uncertainty estimates. To scale this to larger networks, Guo et al. proposed FedEP, an expectation propagation approach where clients iteratively refine a global posterior distribution, yielding significantly better-calibrated models than standard optimization.

Gaussian VI:     [ Mean (μ) , Variance (σ²) ]  <-- Constrained to a rigid shape
Particle VI:     [ Particle 1 , Particle 2 , Particle 3 ] <-- Captures arbitrary, multi-modal shapes

Assuming the posterior fits a neat Gaussian shape can be overly restrictive for deep neural networks. To capture complex, multi-modal uncertainty landscapes, Wei et al. introduced a particle-based framework (FedWBA). Clients maintain a collection of distinct model instances ("particles") that explore the parameter space via Stein Variational Gradient Descent. The server then merges these particle sets using Wasserstein Barycenter Aggregation, preserving non-Gaussian uncertainty characteristics across heterogeneous clients.


Functional Space Collaboration and Distribution-Free Calibration

An alternative strategy shifts the focus entirely away from model weights. Because deep neural networks suffer from weight-space symmetries (where completely different weights yield identical predictions), aggregating weight distributions can be noisy and inefficient. Instead, some frameworks quantify uncertainty directly in the functional (output) space.

  • Functional Bayesian Networks: In FedBNN, developed by Makhija et al., clients do not share weight distributions. Instead, they collaborate by aligning their predictive distributions over a shared set of anchor data points. This allows edge devices to run entirely different model architectures while still maintaining high entropy (high uncertainty) when encountering out-of-distribution data.
  • Personalized Gaussian Processes: Achituve et al. proposed pFedGP, which uses a shared deep neural network across clients as a feature extractor, but deploys personalized Gaussian Processes (GPs) at the edge. This provides mathematically rigorous, closed-form confidence intervals directly at the prediction layer.
  • Federated Conformal Prediction: If you require strict calibration guarantees without making any Bayesian assumptions about the data distribution, you can use conformal prediction. Lu et al. introduced Federated Conformal Prediction (FCP). Instead of outputting a single class guess, FCP computes a prediction set (e.g., predicting that an image contains [Lung Nodule, Cyst]) that is mathematically guaranteed to contain the true label with a user-defined probability, even when data is distributed non-identically across nodes.

Why It Works: The Intuition Behind the Calibration Payoff

To understand why this shift works, we must look at how uncertainty changes the way decentralized models interact.

In standard federated learning, a client that trains on a highly biased local dataset will aggressively update its model weights, and the server will blindly accept those updates. If a hospital model has never seen an image of a spleen, it will still output a high-confidence prediction when forced to evaluate one, corrupting the global model during aggregation.

By introducing uncertainty quantification, the model learns the boundaries of its own knowledge. In a diverse label setting like FUNAvg (Tölle et al.), the server uses a client's local predictive uncertainty as an aggregation weight. If a client's local model exhibits high entropy for a specific class, the server discounts its update for that class.

Ultimately, quantifying uncertainty transforms federated learning from a system that merely averages what every client knows into a system that maps out, respects, and protects against what each client does not know.

Key Foundational Work

Differentially Private Federated Variational Inference (2019) — contributed 2 concept elements to this direction.
Distributed Inference for Linear Support Vector Machine (2018) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_uncertainty_quantification_in_federated__fdee57c4,
  title = {uncertainty quantification in federated learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/uncertainty-quantification-in-federated-learning/rdr_fdee57c417304880a42a9e432a6a30a9},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_uncertainty_quantification_in_federated__fdee57c4_v1,
  title = {uncertainty quantification in federated learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/uncertainty-quantification-in-federated-learning/rdr_fdee57c417304880a42a9e432a6a30a9/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake