Towards Personalized Federated Learning

In plain words

Federated learning trains one shared model across many devices or organisations, and the raw data never leaves each of them. The trouble is that each participant's data can look quite different from the others'. A single shared model then fits the average participant well and fits nobody in particular very well.

This survey looks at the fixes for that problem, under the name personalized federated learning. The fixes fall into two families. One family keeps a single shared model and lets each participant adjust it on its own data, or changes how the shared model is trained so that adjusting it works better. The other family gives each participant its own model, and lets participants with similar data share what they learn. The authors sort the existing methods into these families, list the benchmarks and measures people use to test them, and point to open questions such as new model designs, fairer benchmarks and trust.

The survey is a map of the field rather than a new method. It runs no experiments of its own.

TL;DR

• A survey of personalized federated learning (PFL), the variant where each client gets a model suited to its own non-IID data §I.
• Methods are organised in a taxonomy: global-model personalization (data-based, model-based) and learning personalized models (architecture-based, similarity-based) §II fig. 1.
• Benchmark and metric coverage is summarised in tables; the authors find that most PFL work tests only one type of non-IID skew §V tab. III.
• No new experiments and no new numbers. Use it to find the family of a method, then read the original paper for its formulas.

Problem

Standard federated averaging (FedAvg) trains one global model over clients whose data differ. With non-IID data its accuracy drops, an effect the authors attribute to client drift: local steps pull each client's model away from the global optimum, and the average of those models misses it §I-B fig. 2. Even with good convergence, one global model fits the "average client" and generalises poorly to a client whose distribution is far from the global one §I-B.

The survey frames PFL as the middle ground between two extremes §II. Standard FL shares knowledge but gives no per-client output eq. 1eq. 2. Local learning gives each client a full model but uses no collaboration, so each model sees few examples eq. 3. The survey assumes that the federation is horizontal FL (clients share the feature space and differ in samples) and states that most PFL work targets that setting §I-A.

Method

The survey's organising unit is the taxonomy, not one algorithm. The steps below are the survey's own structure.

Background objectives §II.

1. Standard FL minimises the unweighted average of client losses over one global parameter vector w eq. 1eq. 2. The paper writes the average with a 1/C factor and does not discuss weighting by data size.
2. Local learning gives each client its own parameters θ_c and minimises that client's loss alone eq. 3.

Strategy I: global model personalization §III. Train one global model with FL, then adapt it per client with a local step. The survey notes that a better global model should give better personalization §III.

1. Data-based approaches reduce the heterogeneity of client data §III-A.
   • Data augmentation: share a small balanced dataset, train a GAN on the server, or oversample on the client §III-A.
   • Client selection: choose a subset of clients per round by bandit or Q-learning, or by tier or loss §III-A.
2. Model-based approaches change the local objective or the global model §III-B.
   • Regularized local loss adds a term l_reg to the local loss, built from the global model w and the local model θ eq. 4. Between global and local models the examples are FedProx, FedCL (importance weights from elastic weight consolidation) and SCAFFOLD (a correction term from the difference between global and local update directions) §III-B. Between historical local snapshots, MOON uses a contrastive loss §III-B. The survey gives no formula for FedProx, SCAFFOLD or MOON, so consult the originals.
   • Meta-learning treats clients as tasks. Per-FedAvg minimises the global loss after one local gradient step w - α ∇f_c(w) eq. 5, where α > 0 is the step size. It notes that FO-MAML drops the Hessian term and HF-MAML replaces the Hessian-vector product with gradient differences. pFedMe adds an l2 regularizer; ARUBA uses online learning for the meta step §III-B.
   • Transfer learning reuses the lower layers of the global model and fine-tunes the upper layers locally. Domain adaptation can add a CORAL alignment layer before the softmax §III-B fig. 3. FedMD distils from a model pre-trained on a public dataset §III-B.

Strategy II: learning personalized models §IV. Each client gets its own model, shaped by its architecture or by its similarity to other clients.

1. Architecture-based approaches §IV-A.
   • Parameter decoupling keeps some layers private. The base-plus-personalized design keeps deep layers local and shares the base layers §IV-A. A second design keeps client-specific embeddings private, as in LG-FedAvg §IV-A fig. 4.
   • Knowledge distillation trains clients with different architectures. The survey lists four directions: distil to clients, to the server, both ways, or between clients §IV-A fig. 4.
     • FedMD averages class scores on a public dataset as a consensus, then fine-tunes on private data §IV-A.
     • FedGen trains a generator on the server and clients regularise with generated representations §IV-A.
     • FedDF groups clients into prototype architectures, runs FedAvg per group, then distils ensembles on an unlabelled public dataset §IV-A.
     • FedGKT distils both ways with a KL loss; the survey notes that ground-truth labels are sent to the server §IV-A.
2. Similarity-based approaches §IV-B.
   • Multi-task learning treats each client as a task. MOCHA uses a primal-dual formulation for convex models only. VIRTUAL uses variational inference. FedAMP forms a personalized cloud model u_c as a weighted combination of client models, then solves a proximal local step eq. 6.
   • Model interpolation mixes local and global models. The penalty λ sets the degree of personalization: λ = 0 is pure local training and λ → ∞ forces identical models §IV-B. APFL learns the mixing weight per client §IV-B. HeteroFL gives each client a model of a chosen complexity, all derived from one global model §IV-B.
   • Clustering groups clients with a cluster-specific model §IV-B. The survey contrasts IFCA, multi-center learning with EM, recursive bi-partitioning and hierarchical clustering §IV-B eq. 7.

# Two-step PFL: "FL training + local adaptation"  [§III]
# Input:  clients with local datasets; number of FL rounds
# Output: one personalized model per client

def personalize_global_model(clients, rounds):
    # 1. Train one global model with ordinary FL  [§III] [eq. 1]
    global_model = federated_train(clients, rounds)

    # 2. Adapt the global model on each client's own data  [§III]
    personal_models = {}
    for client in clients:
        personal_models[client] = fine_tune(global_model, client.dataset)

    return personal_models

# Iterative clustering: one model per cluster, lowest loss wins  [§IV-B]
# Input:  K cluster models; clients with local datasets
# Output: updated cluster models; cluster index per client

def clustered_round(cluster_models, clients):
    # 1. Each client scores all K cluster models on its data  [§IV-B]
    for client in clients:
        losses = [loss(model, client.dataset) for model in cluster_models]
        # 2. The client joins the cluster with the lowest loss  [§IV-B]
        client.cluster = argmin(losses)

    # 3. The server averages updates within each cluster  [§IV-B]
    for k in range(K):
        cluster_models[k] = average_within(clients, k)

    return cluster_models

Benchmarks and metrics §V. The survey lists benchmark suites (LEAF, FLBench, Edge AIBench, OARF, FedGraphNN) and real federated datasets, but notes that most PFL papers partition public data themselves §V.

1. Non-IID types used in PFL: quantity skew, feature distribution skew, label distribution skew and label preference skew §V tab. III. A common label-skew recipe draws per-client label proportions from a Dirichlet distribution with parameter α; α = 100 is called IID and smaller α gives more skew §V.
2. Metrics are grouped as model performance, system performance and trustworthy AI §V tab. IV. The survey argues that the average test accuracy of personalized models hides per-client differences and that the cost side (communication, computation, scalability) is rarely reported §V.

Future directions §VI. Architecture design (heterogeneity analytics, aggregation beyond averaging, neural architecture search, spatial and temporal adaptability), benchmarking (realistic datasets, realistic non-IID settings, holistic metrics) and trustworthy PFL (open collaboration and incentives, fairness, explainability, robustness) §VI.

Hyperparameters

The survey runs no training experiments and states no training hyperparameters. The values below are the ones it does state.

Results

No experiments in this survey. It reports no headline number of its own. The figures it quotes come from the works it cites and are author-reported there.

Paper vs code

No official code. The paper has no code repository link. Its only URLs are IEEE template links, and the paper has no project page. I did not run the GitHub search, because the paper is a survey with no implementation. The third-party repositories that list PFL papers are reading lists, not code, so none was used.

Pitfalls

• No experiments. Nothing here can be reproduced or compared. The value is the taxonomy and the pointers into the cited papers.
• Most methods get one sentence. The survey writes formulas only for FedAvg-style objectives eq. 1eq. 2eq. 3, the regularized local loss eq. 4, Per-FedAvg eq. 5, FedAMP eq. 6 and multi-center clustering eq. 7. For FedProx, SCAFFOLD, MOON, pFedMe, APFL, HeteroFL, IFCA and the rest, read the original papers.
• The objective is an unweighted average. Eq. 1 averages client losses with 1/C, and the survey does not discuss weighting by data size. Read FedAvg's original paper before assuming a data-weighted average.
• Equation 3 is written with mixed indices. It minimises over θ_1, ..., θ_C but writes the objective with θ_c inside the sum. Treat it as stacked per-client parameters.
• Per-FedAvg needs second-order gradients through ∇f_c(w) inside the objective eq. 5. The survey notes the cost and the first-order and Hessian-free approximations, but gives no approximation error.
• Multi-center clustering divides by the cluster size in the M-step, and the text does not say what happens if a cluster has no clients eq. 7. An implementer must handle that case.
• Non-IID simulation depends on α. The Dirichlet convention and the α = 100 IID reading come from the cited simulation literature §V. The survey does not justify them against real federated data.
• Coverage gaps the survey itself names. Label preference skew is studied only by clustering methods, and trustworthy-AI metrics are rarely used §V tab. III tab. IV.
• Date. The survey was written in 2021 and cites work up to that point; the arXiv record shows version 3 (2022). Later PFL methods are not covered.
• Numbering. Sections and tables are cited as the paper prints them (Roman numerals, for example §III-B and tab. III). Figures are numbered in order of appearance.

Lineage

• Builds on: the FedAvg algorithm, which most FL methods derive from §I; the paper does not give an arXiv id for it in the text.
• Surveys it sits beside: general FL surveys §I-C, privacy surveys and robustness surveys §I-C. The survey frames itself as the first systematic PFL review for new researchers §I-C.
• Compared against: not applicable (no experiments).
• Superseded by: not stated in the source.