Home / Curated / Algorithms / HarmoFL: Harmonizing Local and Global Drifts in Federated Learning on Heterogeneous Medical Images

HarmoFL: Harmonizing Local and Global Drifts in Federated Learning on Heterogeneous Medical Images

Algorithms AIMech:HarmoFlHarmonizingLocalAndGlobalDriftsInFederatedLearning DRAFT

Aliases: HarmoFL, Harmonizing Local and Global Drifts in Federated Learning, Amplitude Normalization and Client Weight Perturbation Federated Learning

A federated learning algorithm for feature-heterogeneous multi-institution imaging (HarmoFL; Jiang, Wang, and Dou, AAAI 2022) that jointly mitigates local client drift and global server drift without extra model-parameter communication overhead by combining (1) frequency-domain Amplitude Normalization—where each client i updates a moving-average Fourier amplitude spectrum A_bar_{i,k} = (1 - v) A_bar_{i,k-1} + v (1/M) sum_{m=1}^M A_{i,x_m} (initialized each round from the server-averaged global amplitude A_bar^t = (1/N) sum_{i=1}^N A_bar_{i,K}^{t-1}) and reconstructs harmonized images Psi(x_m) = F^{-1}(A_bar_{i,k}, P_{i,x_m}) preserving local phase P_{i,x_m}—with (2) Client Weight Perturbation, where each local step computes Euclidean-normalized gradient ascent perturbation delta_k = alpha * nabla ell_i(Psi(x), y; theta_{i,k-1}) / ||nabla ell_i(Psi(x), y; theta_{i,k-1})||_2 and updates theta_{i,k} = theta_{i,k-1} - eta_l nabla ell_i(Psi(x), y; theta_{i,k-1} + delta_k) to guide local models into overlapping flat optima prior to server aggregation.

Identity Notes & Concept Scope Distinct from FedDG / ELCFS (AIMech:FedDgEpisodicLearningInContinuousFrequencySpaceForFederatedDomainGeneralization; Liu et al., CVPR 2021), which shares per-client frequency amplitude banks for episodic continuous interpolation and boundary-oriented meta-learning; HarmoFL maintains a single server-averaged moving-average amplitude template A_bar^t without sharing individual image amplitude banks and couples frequency harmonization with SAM-style local Euclidean weight perturbation delta_k = alpha * g / ||g||_2. Distinct from FedBN (Li et al., ICLR 2021), which keeps Batch Normalization parameters local without frequency-space amplitude harmonization or local weight perturbation.

Claims & Grounded Evidence

4 Claims

Mechanism #harmofl-amplitude-normalization-and-weight-perturbation-mechanism

In each communication round t, client i initializes its running amplitude A_bar_{i,0}^t = A_bar^t from the server-averaged amplitude spectrum, updates A_bar_{i,k}^t = (1 - v) A_bar_{i,k-1}^t + v (1/M) sum_{m=1}^M A_{i,x_m} at batch k (with v = 1 in round 1 and v = 0.1 thereafter), reconstructs amplitude-normalized inputs Psi(x_m) = F^{-1}(A_bar_{i,k}^t, P_{i,x_m}), and performs local gradient descent evaluated at the perturbed weights theta_{i,k-1}^t + delta_k where delta_k = alpha * nabla ell_i(Psi(x), y; theta_{i,k-1}^t) / ||nabla ell_i(Psi(x), y; theta_{i,k-1}^t)||_2, after which the server averages both model weights theta^{t+1} = theta^t - eta_g (1/N) sum_{i=1}^N Delta theta_i^t and amplitude spectra A_bar^{t+1} = (1/N) sum_{i=1}^N A_bar_{i,K}^t.

supports Source: HarmoFL: Harmonizing Local and Global Drifts in Federated Learning on Heterogeneous Medical Images · Locator: Sections 3.1-3.2, Eqs. (1)-(8), Algorithm 1, pp. 3-4

Scope #harmofl-coupled-drift-bound-and-nonconvex-convergence-rate

Under L-smoothness, bounded gradient norm G, bounded Hessian norm B, bounded parameter norm C, and bounded cross-client gradient difference (1/N) sum_{i=1}^N ||nabla ell_i(Psi(x), y; theta) - nabla f(theta)|| <= epsilon (Assumptions B.1-B.4), Theorem B.6 shows that overall client-server drift Gamma = (1/(KN)) sum_{k=1}^K sum_{i=1}^N E[||theta_{i,k}^t - theta^t||^2] satisfies Gamma <= 4 eta_tilde^2 (G^2 + B^2 C^2) / eta_g^2 + 4 eta_tilde^2 epsilon^2 (N-1)^2 / (eta_g^2 N^2) + 2 eta_tilde^2 sigma^2 / (K eta_g^2) with eta_tilde = K eta_g eta_l, and Theorem 3.1 proves non-convex stationarity convergence at rate O(sqrt((f(theta^0) - f(theta^*)) / (K T)) + (f(theta^0) - f(theta^*)) / T).

supports Source: HarmoFL: Harmonizing Local and Global Drifts in Federated Learning on Heterogeneous Medical Images · Locator: Section 3.3 Theoretical Analysis, Theorem 3.1, Eq. (9), p. 5; Appendix B Assumptions B.1-B.4, Theorem B.5, Theorem B.6, pp. 9-11

Limitation #harmofl-double-gradient-evaluation-and-perturbation-radius-limitation

Computing the perturbed gradient nabla ell_i(Psi(x), y; theta_{i,k-1} + delta_k) requires two forward-backward passes per local step (first at theta_{i,k-1} to form delta_k and second at theta_{i,k-1} + delta_k) Plus 2D FFT/IFFT operations on each image batch, and the theoretical convergence guarantee requires the perturbation radius alpha to decay as O(1 / sqrt(K T)) because overly large alpha degrades local optimization stability (Appendix C.2 Figure 6, p. 12).

supports Source: HarmoFL: Harmonizing Local and Global Drifts in Federated Learning on Heterogeneous Medical Images · Locator: Section 3.2 Eqs. (6)-(7), p. 4; Section 3.3 Theorem 3.1, p. 5; Appendix C.2 Figure 6, p. 12

Application #harmofl-multi-site-medical-image-classification-and-segmentation-application

Applied to multi-institution federated medical image classification and segmentation across scanner and staining protocol shifts, including histology nuclei segmentation, prostate MRI segmentation (6 institutions), and Camelyon17 breast cancer lymph-node histology classification (5 hospitals).

supports Source: HarmoFL: Harmonizing Local and Global Drifts in Federated Learning on Heterogeneous Medical Images · Locator: Section 4 Experiments, Tables 1-2, Figures 3-5, pp. 5-8

⚠️ Curated Limitations & Negative Results

Client weight perturbation requires two gradient computations per local mini-batch step (evaluating at theta_{i,k-1} and at theta_{i,k-1} + delta_k), roughly doubling local backpropagation cost alongside 2D FFT/IFFT image transformations (Section 3.2, pp. 3-4).
Sharing the moving-average amplitude spectrum A_bar_{i,K}^t with the central server assumes that low-level intensity and contrast shifts are captured primarily in Fourier amplitudes while anatomical semantics reside in phases, which is tailored to multi-scanner imaging shifts rather than non-spectral tabular or text modalities (Section 3.1, p. 3).
Convergence guarantees (Theorem 3.1) require the perturbation radius alpha to decay as O(1 / sqrt(K T)), and overly large fixed alpha values degrade accuracy (Section 3.3, p. 5; Appendix C.2 Figure 6, p. 12).

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:FederatedLearning` Federated Learning	broad_match	HarmoFL is a specialized federated learning algorithm combining frequency-space amplitude normalization and local weight perturbation.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`r501-primary-pdf`	HarmoFL: Harmonizing Local and Global Drifts in Federated Learning on Heterogeneous Medical Images Meirui Jiang, Zirui Wang, Qi Dou	primary paper full_text	ID: https://doi.org/10.1609/aaai.v36i1.19993 sha256:116b5072a165…
`r501-arxiv-abs`	HarmoFL: Harmonizing Local and Global Drifts in Federated Learning on Heterogeneous Medical Images (arXiv abstract) Meirui Jiang, Zirui Wang, Qi Dou	primary paper abstract_only	ID: https://arxiv.org/abs/2112.10775 sha256:1426048d7e76…
`r501-metadata-cr`	HarmoFL: Harmonizing Local and Global Drifts in Federated Learning on Heterogeneous Medical Images (Crossref metadata) Meirui Jiang, Zirui Wang, Qi Dou	primary paper abstract_only	ID: https://doi.org/10.1609/aaai.v36i1.19993 sha256:bce74fe7f1f0…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini

Created: 2026-10-19T09:44:00Z

Updated: 2026-10-19T09:44:00Z

Source File: data/curated/HarmoFlHarmonizingLocalAndGlobalDriftsInFederatedLearning.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