Home / Curated / Algorithms / FedPer (Federated Learning with Personalization Layers)

FedPer (Federated Learning with Personalization Layers)

Algorithms AIMech:FedPerFederatedLearningWithPersonalizationLayers DRAFT

Aliases: FedPer, Federated Learning with Personalization Layers, Base Plus Personalization Layer Federated Learning

A personalized federated deep learning algorithm (Arivazhagan, Aggarwal, Singh, & Choudhary 2019, Section 3, Equations 1-3, Algorithms 1-2) that decomposes each client j's deep feedforward network into K_B shared base layers W_B = (W_{B,K_B}, ..., W_{B,1}) and K_P client-specific personalization layers W_{P_j} = (W_{P_j,K_P}, ..., W_{P_j,1}) producing predictions y_hat = f(x; W_B, W_{P_j}) (Equation 1), minimizes the average personalized population risk L^{PR}(W_B, W_{P_1}, ..., W_{P_N}) = (1/N) sum_{j=1}^N E_{(x,y)~P_j}[l(y, f(x; W_B, W_{P_j}))] (Equation 2) by having each client j jointly update both its base and personalization layers locally via e epochs of stochastic gradient descent (W_{B,j}^{(k)}, W_{P_j}^{(k)}) <- SGD_j(W_B^{(k-1)}, W_{P_j}^{(k-1)}, eta_j^{(k)}) (Algorithm 1, Line 5), retaining W_{P_j}^{(k)} privately on device j while transmitting only W_{B,j}^{(k)} to the parameter server for sample-weighted FedAvg aggregation W_B^{(k)} <- sum_{j=1}^N gamma_j W_{B,j}^{(k)} with gamma_j = n_j / sum_{l=1}^N n_l (Algorithm 2, Line 6).

Identity Notes & Concept Scope Distinct from FedRep (AIMech:FedRepExploitingSharedRepresentationsForPersonalizedFederatedLearning, Collins et al. 2021), which alternates separate inner head-only optimization epochs and representation-only optimization steps within each round rather than jointly updating (W_{B,j}, W_{P_j}) simultaneously via coupled SGD steps (Algorithm 1, Line 5); and distinct from FedBABU (AIMech:FedBabuFederatedLearningWithOnlyBodyAggregationAndBodyUpdate, Oh et al. 2022), which freezes a shared random head (alpha_h = 0) during federated representation training.

Claims & Grounded Evidence

4 Claims

Mechanism #r473-coupled-local-sgd-base-only-server-aggregation-mechanism

In each global round k, FedPer-Client(j) initializes from the server's aggregated base weights W_B^{(k-1)} and its own private personalization weights W_{P_j}^{(k-1)}, executes e local SGD epochs jointly over (W_B, W_{P_j}) on empirical risk L_j^{ER}(W_B, W_P) (Equation 3), and uploads only W_{B,j}^{(k)} for gamma_j-weighted server averaging while keeping W_{P_j}^{(k)} strictly private on client j (Arivazhagan et al. 2019, Section 3.2, Algorithms 1-2).

supports Source: Federated Learning with Personalization Layers · Locator: Section 3.2 (Algorithms 1-2), p. 4

Scope #r473-contradictory-user-preference-and-reduction-to-fedavg-scope

FedPer reduces exactly to standard FedAvg when K_P = 0, while for K_P >= 1 the private top layers W_{P_j} enable different clients to assign distinct or contradictory personalized labels to identical input features (where single-model FedAvg degrades to near-random guessing) and reduce cross-client test accuracy variance under non-identical class partitions (Arivazhagan et al. 2019, Sections 1, 4.2, & 4.5, Figures 2-3 & 7-8).

supports Source: Federated Learning with Personalization Layers · Locator: Sections 4.2-4.5 (Figures 2-8), pp. 5-8

Application #r473-personalized-image-aesthetics-and-non-iid-vision-application

Applied to personalized image aesthetics rating prediction on the FLICKR-AES dataset and non-identically partitioned image classification on CIFAR-10 and CIFAR-100 with ResNet-34 and MobileNet-v1 architectures (Arivazhagan et al. 2019, Section 4, Figures 2-18).

supports Source: Federated Learning with Personalization Layers · Locator: Section 4.1-4.5, pp. 5-8, and Appendix A, pp. 10-13

Limitation #r473-base-personalization-desynchronization-between-rounds-limitation

When the server replaces a client's local base layers W_{B,j}^{(k-1)} with the newly aggregated global base W_B^{(k-1)} at the start of round k, the private personalization layers W_{P_j}^{(k-1)} can be temporarily mismatched to the updated base representation until local fine-tuning or joint SGD epochs re-align them (Arivazhagan et al. 2019, Appendix A.3, Figures 15-16).

supports Source: Federated Learning with Personalization Layers · Locator: Appendix A.3 (Effect of FineTuning, Figures 15-16), pp. 10-12

⚠️ Curated Limitations & Negative Results

Requires stateful clients that persist their private personalization layers W_{P_j} across communication rounds and cannot directly classify on unseen stateless clients without local labeled samples.
Requires all participating clients to share an identical architecture for the K_B base layers W_B.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:FederatedLearning` Federated Learning	broad_match	FedPer is a personalized federated deep learning algorithm within the broader FederatedLearning category.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`r473-primary-pdf`	Federated Learning with Personalization Layers Manoj Ghuhan Arivazhagan, Vinay Aggarwal, Aaditya Kumar Singh, Sunav Choudhary	primary paper full_text	ID: https://arxiv.org/abs/1912.00818 sha256:ec6546bd9325…
`r473-arxiv-abs`	Federated Learning with Personalization Layers (arXiv abstract) Manoj Ghuhan Arivazhagan, Vinay Aggarwal, Aaditya Kumar Singh, Sunav Choudhary	primary paper abstract_only	ID: https://arxiv.org/abs/1912.00818 sha256:1bdfe17cbfdb…
`r473-metadata-cr`	Federated Learning with Personalization Layers (Crossref metadata) Manoj Ghuhan Arivazhagan, Vinay Aggarwal, Aaditya Kumar Singh, Sunav Choudhary	primary paper abstract_only	ID: https://arxiv.org/abs/1912.00818 sha256:c8da39fce095…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini

Created: 2026-10-19T09:16:00Z

Updated: 2026-10-19T09:16:00Z

Source File: data/curated/FedPerFederatedLearningWithPersonalizationLayers.yaml · Schema: curated_concept.yaml

Explore across X-Mech

Related computational & biological Mechs: Embedding Map Network Graph ProteinTraitsMech (ESM-2 pLM) DUFMech (Domain pLM) TraitMech CultureMech All 14 X-Mech Repositories