Source-linked AI summary
Beyond Non-IID: Learner--Client Distribution Mismatch in Federated Learning
Yiming Xie, Lili Su, Ningfang Mi
TL;DR
The paper addresses federated learning when participating clients’ distributions mismatch a learner’s target distribution, making uniform client aggregation unreliable. It proposes DIC-KT, which uses a small target proxy set to estimate clients’ marginal utility dynamically; on CIFAR-10, it outperforms static and dynamic baselines, especially under severe mismatch.
Problem
Learner-centric federated learning must optimize a specific target distribution that may differ substantially from aggregate client data, while existing methods often use static utility signals or centralized, one-shot source assessment.
Method
DIC-KT dynamically estimates each client’s marginal contribution using a small target proxy set and jointly optimizes client selection and aggregation.
Results
DIC-KT consistently outperforms static and dynamic baselines on CIFAR-10, achieving higher accuracy with lower variance, particularly under severe distribution mismatch.
Takeaways & Limitations
Effective collaboration under learner-client mismatch is a knowledge-transfer problem requiring adaptive identification and prioritization of clients’ evolving contributions.
Takeaways & Limitations
Existing methods remain limited by static usefulness assumptions and cannot adapt to clients’ evolving contributions during training.
Abstract
from arXiv · showhide
Federated learning systems are increasingly deployed to facilitate collaborative model training across a heterogeneous client population. Existing practice mostly implicitly assumes that the aggregated client data distribution is representative of the learner's target distribution or that learning from all available clients is uniformly beneficial for the learner distribution. However, such an assumption often does not hold in reality. Traditional client selection strategies in FL literature largely overlook such misalignment, while most existing work on multi-source transfer learning either requires direct access to local data or uses one-shot model/feature aggregation. In this paper, we take the initiative to understand and mitigate the impacts of such learner-client population misalignment. In particular, we consider the practical setting where the learner keeps a small proxy dataset. We observe that client contributions vary significantly across training rounds, and traditional technology is insufficient to identify beneficial sources under multi-source transfer diversity. Then, we propose a dynamic, influence-aware client selection framework that estimates each client's potential utility to the learner's optimization objective using proxy influence signals on a learner-specific proxy set. Via using leave-one-out evaluations, we prioritize the most informative sources of knowledge while controlling the negative impacts of statistical noise and data heterogeneity. Experiments on CIFAR-10 under heterogeneous data partitions demonstrate that our approach consistently outperforms static and dynamic baselines, achieving faster convergence and higher accuracy.
I. INTRODUCTION
Federated learning can misalign with a learner’s target distribution because participating clients may not represent deployment users. The paper frames this as learner-centric multi-source transfer and proposes influence-aware client selection using proxy data.
- Motivation: Classical FL minimizes average client loss, implicitly treating all clients symmetrically despite a learner-specific target objective.The learner may instead seek performance on a specific target distribution.
- Motivation: Client and target distributions can differ substantially, so using every available client may dilute target-relevant information and induce negative transfer.The paper highlights narrow opt-in demographics serving broader or different user populations as one practical example.
- Problem Formulation: Learner-centric FL is formulated as multi-source transfer in which heterogeneous client contributions vary in distributional alignment and quantity.Both source quality and contribution amount affect transfer performance.
- Approach: DIC-KT estimates each client’s marginal contribution to the learner objective with a small learner-specific proxy dataset, without accessing raw client data.Leave-one-out evaluations prioritize informative sources while controlling statistical noise and heterogeneity.
- Limitations of Existing Methods: The paper identifies static or traditional dynamic selection as insufficient under distribution mismatch and limited learner data.These strategies do not explicitly account for evolving learner-centric utility.
- Results: Experiments demonstrate consistent empirical improvements across heterogeneous and misaligned data partitions, supporting adaptive source selection for knowledge transfer.The reported gains include improved convergence and accuracy relative to static and dynamic baselines.
B. Multi-Source Transfer Learning
Multi-source transfer learning studies how heterogeneous auxiliary sources can improve a target task, but source relevance and quantity vary, making indiscriminate transfer risky. Existing approaches often require centralized source access or assess utility only once, motivating learner-centric adaptive selection.
- Negative transfer can occur when poorly aligned sources are transferred indiscriminately to a target task.Source-target divergence and source relevance affect transfer performance.
- Source usefulness depends on both distributional relevance and the amount of transferred knowledge.Transferring too much knowledge from a relevant source can also be suboptimal.
- Most multi-source methods assume centralized access to source datasets or models and often use one-shot utility assessment.H-Ensemble avoids raw-source disclosure but fuses source feature extractors and overlooks later knowledge refinement.
- Federated client-selection methods use system, gradient, loss, or contribution signals, but many target throughput or global validation objectives rather than learner-specific utility.Validation-based marginal-effect methods can also incur substantial computational overhead.
- DIC-KT addresses learner-centric selection by adapting client choices to evolving potential contributions under target-client distribution mismatch.Clients are treated as external knowledge sources whose utility varies with relevance and transferability.
A. Challenges and Limitations of Existing Methods
Existing selection strategies do not explicitly model learner-centric utility under target-client mismatch, especially when heterogeneous sources change usefulness during training. Controlled CIFAR-10 cases show that similarity signals can be weak in homogeneous or overlapping settings, although they help when structural heterogeneity is clear.
- A. Challenges and Limitations of Existing Methods: Existing methods largely rely on static similarity or proxy optimization signals and do not explicitly account for learner-centric utility under mismatch.Their usefulness is treated as largely static, preventing adaptation to evolving client contributions.
- A. Challenges and Limitations of Existing Methods: Uniform sampling and static similarity can produce inefficient training and unstable convergence when client heterogeneity is high and transfer utility evolves.This motivates adaptive mechanisms that directly evaluate client-update impact on the learner objective.
- B. Empirical Insights under Multi-Source Transfer Diversity: The four CIFAR-10 cases vary learner label support and inter-client heterogeneity to evaluate alignment-aware selection under controlled regimes.Cases include homogeneous clients, restricted learner supports, disjoint match and exclude groups, and Dirichlet heterogeneity with α = 0.1.
- B. Empirical Insights under Multi-Source Transfer Diversity: In homogeneous cases, JSD provides limited discrimination and repeatedly selects a small client subset because client label distributions are nearly identical.Random sampling explores a broader range of clients in the corresponding selection plots.
- B. Empirical Insights under Multi-Source Transfer Diversity: In Case 3, JSD increasingly focuses on learner-label-matched clients, whereas Random samples from both groups without distributional relevance.Clear structural heterogeneity makes distribution similarity an effective selection signal.
- B. Empirical Insights under Multi-Source Transfer Diversity: With partially overlapping client distributions, similarity alone provides weaker discrimination than in the structurally separated case.Case 4 uses Dirichlet sampling with α = 0.1, producing skewed and diverse class proportions without strict label separation.
IV. FEDERATED LEARNING WITH INFLUENCE FUNCTIONS
The problem setting separates a learner with a small target-distribution proxy dataset from heterogeneous clients that share only model updates. The learner therefore evaluates client utility against its own target risk rather than the aggregate client objective, and DIC-KT uses influence-based selection for this purpose.
- DIC-KT is a learner-centric client-selection algorithm based on an influence function.The algorithm is formally described as the paper’s proposed selection procedure.
- The system contains one learner and C distributed clients with private datasets, while raw client data are never shared.Client distributions may differ arbitrarily from one another and from the learner target distribution.
- The learner maintains a small proxy dataset sampled from its target distribution and uses it to assess client-update relevance.The learner does not perform local training or generate model updates.
- The setting makes no assumptions about client similarity, clustering, or parametric distribution structure, and clients are accessible only through exchanged model updates.This permits substantial statistical heterogeneity without requiring a small number of client groups.
- Client-objective minimization does not generally minimize learner risk under distribution mismatch, especially when the learner dataset is small.The learner’s proxy dataset serves as an approximation because the target distribution is unknown to participants.
B. Influence-Based Client Contribution Estimation
The framework estimates each client’s round-specific contribution by comparing aggregated and leave-one-out models on the learner’s proxy set, then converts these utility signals into influence-aware aggregation weights.
- Influence estimation: Leave-one-out evaluation measures each client’s marginal effect by comparing the full aggregate with an aggregate that excludes that client.Both models are evaluated on the learner proxy set using renormalized sample-size weights.
- Weighting: The weighting rule uses η to control prioritization sensitivity and ϵ to stabilize weights for clients with nonpositive utility.Larger η amplifies differences between high- and low-utility clients.
- Utility interpretation: A positive utility delta means removing a client increases learner loss, indicating that the client provides useful knowledge.Such clients receive larger influence weights after applying the weighting rule.
- Utility interpretation: Nonpositive utility indicates that a client is redundant or potentially harmful because removing it does not increase learner loss.These clients are down-weighted rather than completely removed.
C. Knowledge Transfer Gain and Learner-Centric Utility
The paper evaluates federated knowledge transfer from the learner’s perspective by contrasting local-only and influence-aware transferred models through a learner-centric gain defined on the target distribution.
- Learner-centric utility: The learner-centric utility analysis evaluates whether client updates reduce the learner’s underlying generalization loss.Leave-one-out evaluation on the learner proxy set estimates this marginal utility without inspecting raw client data.
- Model comparisons: The no-transfer baseline is a model trained solely on the learner’s local dataset S0 using empirical risk minimization.Its parameters are optimized with stochastic gradient descent.
- Model comparisons: The transferred learner model is produced after T rounds of influence-aware federated training and integrates prioritized updates from multiple clients.The model corresponds to the final output of Algorithm 1.
- Knowledge transfer gain: Knowledge transfer gain quantifies the benefit of incorporating external knowledge, with G > 0 indicating successful transfer of useful knowledge.In practice, target-distribution risks are estimated using the learner’s dataset S0 because D0 is unknown.
V. EVALUATION RESULTS
The evaluation implements DIC-KT and representative baselines under a common federated training protocol in the Flower framework.
- Evaluation protocol: DIC-KT and representative baselines are evaluated in Flower using common rounds of client selection, local optimization, update transmission, and central aggregation.The passage describes the evaluation protocol but reports no numerical outcome.
A. Model and Training Settings
The experiments use CIFAR-10 image classification with client data partitioned homogeneously or heterogeneously and a CNN model trained in the federated setting.
- Dataset: CIFAR-10 provides 60,000 color images across 10 classes, including 50,000 training and 10,000 test samples.The dataset supports both homogeneous (IID) and heterogeneous (non-IID) client partitions.
- Model: The model is a CNN with three convolutional blocks followed by a two-layer fully connected classifier head.The passage states that all models are trained under the described experimental setting.
B. Federated Data Partition and Learner Setup
The federation uses strongly heterogeneous CIFAR-10 client partitions, while a fixed learner target set and small proxy set support learner-centric evaluation and feedback.
- Twenty auxiliary clients and one designated learner form the simulated federation.CIFAR-10 is partitioned across clients using a Dirichlet distribution.
- The Dirichlet concentration is α = 0.1, inducing strong label heterogeneity across clients.Smaller α values correspond to more severe label skew.
- The learner target set contains 2,000 fixed samples drawn with α = 0.1 and defines the objective distribution.The same target set is used across methods and communication rounds.
- A small learner proxy dataset with the target set’s label support provides learner-side feedback for applicable methods.
C. Baselines and Accuracy Comparisons
The study compares DIC-KT with selection methods based on random sampling, similarity, distributional matching, ensembles, and training loss. Under the stated setup, DIC-KT achieves the strongest accuracy and stability.
- Baselines: Five baselines represent uniform random, clustering-based, divergence-based, ensemble-based, and high-loss client selection.The methods are FedAvg, KMeans, FL+HC, H-Ensemble, and Power-of-Choice.
- Experimental comparison: Fig. 5 reports average learner test accuracy every 25 communication rounds with α = 0.1 and selection ratio s = 0.25.The ratio selects five clients per round.
- Selection behavior: Fig. 6 visualizes selected client IDs across communication rounds for DIC-KT and all baseline strategies.Each marker denotes a selected client at a given round.
- Results: DIC-KT consistently achieves the highest learner test accuracy and a more stable convergence trajectory across communication rounds.The comparison includes FedAvg, FL+HC, KMeans, H-Ensemble, and Power-of-Choice.
D. Sensitivity Analysis and Discussion
Sensitivity analyses show that DIC-KT is particularly effective under severe mismatch and remains robust across label alignment, heterogeneity, selection ratio, and proxy-set size conditions.
- Label-exclusive case: In the label-exclusive setting, DIC-KT achieves the highest accuracy across all selection ratios.It identifies and exploits the small subset of label-aligned clients while suppressing negative transfer from irrelevant sources.
- Mismatch levels: At α = 0.1, DIC-KT improves accuracy over FedAvg by approximately 35.9%, KMeans by 37.4%, Power-of-Choice by 78.4%, FL+HC by 15.2%, and H-Ensemble by 7.5%.It also exhibits substantially lower variance under extreme mismatch.
- Mismatch levels: As heterogeneity decreases, performance gaps narrow; at α = 2.0, DIC-KT remains within 1.5% of the best-performing baseline.At α = 0.5, it improves over FL+HC by roughly 2.4% and performs on par with H-Ensemble.
- Selection ratio sensitivity: DIC-KT maintains consistently strong performance across all client selection ratios.The results favor dynamically prioritizing informative clients over indiscriminately increasing participation.
- Learner proxy size sensitivity: A proxy set of 50 samples substantially improves accuracy over 10 samples and quickly approaches the performance of larger sets.Further increases to 500, 1000, or 2000 samples yield only marginal gains.
- Discussion: Overall, DIC-KT is most advantageous under severe distribution mismatch while remaining competitive as client and learner distributions become aligned.The framework dynamically estimates each client’s marginal contribution to the learner’s objective.