PaperlayerView exact revision on arXiv

Source-linked AI summary

Fed-Equilibrium Framework for Topological Pareto Control in Robust and Fair Clinical Federated Learning

Ting Xu, Henry Leung

arXiv:2609.11937v1cs.LGcs.DC

TL;DR

In imbalanced clinical federated networks, geometric defenses do not resolve the dominance of high-volume hubs over minority spokes. Fed-Equilibrium combines geometric filtering with Pareto-based weighting, and reports improved minority convergence while maintaining comparable majority loss.

  • Problem

    In non-IID clinical networks, high-magnitude hub updates can overwhelm minority contributions, and standard aggregation can overlook rare-disease knowledge even after security filtering.

    Source: §I · p. 1

  • Method

    Fed-Equilibrium uses geometric quality assurance to filter gradients, then applies Topological Pareto Control within the verified trust region with weighting selected at the Pareto knee.

    Source: §I · p. 1, C. Layer 3: Synergistic gradient… · p. 2, 2) Stage II: Topological Pareto… · p. 3

  • Results

    At w = 2.0, minority loss fell from 0.857 to 0.340 while majority loss remained comparable (0.981 vs. 0.985).

    Source: B. Empirical identification of the… · p. 4

  • Takeaways & Limitations

    The authors conclude that global generalizability can coexist with local clinical representation while accommodating underrepresented spokes without degrading majority utility.

    Source: §V · p. 6

    Validation was under cooperative conditions; robustness boundaries still require stress-testing against adaptive Byzantine attacks.

    Source: B. Limitations and future directions · p. 6

Abstract

from arXiv · show

The deployment of Federated Learning (FL) in multi-center clinical networks faces the challenge of "knowledge dominance," where high-volume hubs naturally overwhelm minority community nodes, implicitly treating the distinct clinical patterns of smaller cohorts as outliers. Existing geometric defenses provide a security baseline but leave this efficiency-fairness dilemma unresolved. To bridge this gap, we propose Fed-Equilibrium, a framework that advances the paradigm from simple defense to topological equilibrium. Unlike traditional aggregators, Fed-Equilibrium implements a sequential architectural synergy. It utilizes a two-stage gradient control cascade: Stage I (geometric quality assurance) enforces directional consistency via a cosine similarity funnel to filter malicious noise, creating a stabilized manifold; Stage II (topological Pareto control) then actively modulates verified contributions by identifying the optimal Pareto knee point. We validated this framework on a bi-national simulation integrating Canadian (CNODES) and U.S. (SyntheticMass) registries. Experimental results demonstrate that the system simultaneously secures the network against adversarial divergence while accommodating underrepresented signals. Notably, the minority U.S. spoke (representing less than 3% of data volume) achieved deep convergence comparable to the data-rich Canadian hub. This confirms that Fed-Equilibrium effectively counters "knowledge dominance," establishing a true "knowledge commons" where global generalizability does not come at the cost of local clinical representation.

I. INTRODUCTION

Clinical federated networks face a knowledge-dominance problem: security filters do not prevent large hubs from overshadowing minority clinics during aggregation. Fed-Equilibrium combines geometric defense with topological control to address that imbalance.

I. INTRODUCTION

  • Geometric quality assurance filters adversarial divergence through directional consistency, but security alone does not guarantee clinical fairness.The GQA framework uses a cosine similarity funnel,.Source: §I · p. 1
  • Large hubs can overwhelm minority updates during aggregation, leaving rare-disease knowledge overlooked despite security filtering.The resulting model may converge toward the “average” patient rather than capture specialized knowledge in the long tail.Source: §I · p. 1
  • Fed-Equilibrium pairs GQA security with a second stage that actively modulates the remaining benign gradients.The proposed sequence is security first, then equilibrium control,.Source: §I · p. 1
  • TPC identifies a Pareto knee point to give minority spokes leverage against large hubs’ convergence momentum.The framework also incorporates FedQLoRA to learn and preserve underrepresented clinical patterns,.Source: §I · pp. 1–2
  • A bi-national simulation using Canadian CNODES and U.S. SyntheticMass registries is reported to neutralize knowledge dominance.The authors frame this as moving from network safety to fairness,.Source: §I · p. 2

II. METHODS

The methods organize the framework into semantic harmonization, a federated PEFT network, and a gradient-control cascade. The harmonization layer converts heterogeneous registry data into a shared instruction-tuning format.

II. METHODS

  • Fed-Equilibrium is organized as semantic harmonization, a federated PEFT network, and a synergistic gradient-control cascade.The network is modeled as a heterogeneous information network.Source: §II · p. 2

A. Layer 1: Data structuring and harmonization

  • The harmonization layer transforms heterogeneous clinical registries into a unified instruction-tuning format.Fig. 1 depicts this data-harmonization pipeline.Source: A. Layer 1: Data structuring and… · p. 2, Figure 1 · p. 2
  • The simulation assumes unstructured narratives have already been preprocessed through local ETL and NLP, using standardized datasets as post-extraction proxies.This lets the study focus on systematic heterogeneity.Source: A. Layer 1: Data structuring and… · p. 2
  • The pipeline maps Canadian ATC codes and U.S. free-text clinical data into a shared semantic space, then turns harmonized profiles into supervised-finetuning prompts.Each example pairs a synthesized patient context with a target clinical decision or outcome.Source: A. Layer 1: Data structuring and… · p. 2

B. Layer 2: The federated PEFT architecture (FedQLoRA)

FedQLoRA enables collaborative training while keeping raw clinical data at local sites. It reduces communication overhead by training low-rank adapters instead of the full model.

B. Layer 2: The federated PEFT architecture (FedQLoRA)

  • FedQLoRA follows a “bring the model to the data” strategy, keeping raw clinical data within each local environment.The approach supports collaborative learning across diverse sources without data centralization.Source: B. Layer 2: The federated PEFT… · p. 2
  • Each node freezes the TinyLlama-1.1B backbone and optimizes only adapter matrices A and B, whose low-rank decomposition is governed by rank r.This reduces communication overhead relative to full fine-tuning.Source: B. Layer 2: The federated PEFT… · p. 2
  • Each client’s local update is a sparse, efficient gradient vector that reduces bandwidth usage compared with full fine-tuning.The low-rank update’s contribution is controlled by scaling factor α.Source: B. Layer 2: The federated PEFT… · p. 2

C. Layer 3: Synergistic gradient control cascade (TPC)

Fed-Equilibrium combines geometric screening with topological Pareto control: it filters misaligned updates before weighting verified contributions to balance majority and minority influence.

C. Layer 3: Synergistic gradient control cascade (TPC)

  • Stage II operates within the trust region established by Stage I, making the cascade sequential rather than treating security and utility as separate objectives.The paper defines “topological” as control of influence structure over the Pareto manifold, not graph connectivity.Source: C. Layer 3: Synergistic gradient… · p. 2
  • Stage I uses directional alignment and a threshold τ to retain constructive heterogeneity while filtering updates with orthogonal or negative alignment.The gatekeeper stabilizes a trust region for subsequent fairness adjustments.Source: 1) Stage I: GQA gatekeeper · p. 3
  • Stage II weights verified contributions using λk, operationalized as the minority weight wus, selected through a sweep to locate the Pareto knee point.The sweep covers wus ∈ {0.1, . . . , 14.0}.Source: 2) Stage II: Topological Pareto… · p. 3
  • TPC uses the knee point to counterbalance majority convergence momentum and preserve minority insights without compromising convergence.The paper describes the GQA angular bounds as θmin and θmax, the minimum and maximum acceptable deviations from the consensus direction.Source: 2) Stage II: Topological Pareto… · p. 3

D. Simulation and experiment setup

The experiments simulate a 13-node clinical network with a strong data-volume imbalance and compare the proposed cooperative mode with two aggregation baselines.

D. Simulation and experiment setup

  • The simulated network has 10 Canadian hub sites and 3 U.S. minority sites, with approximately 40:1 record imbalance: 37,287 Canadian records versus 974 U.S. records.The network is described as a star topology and simulated in PyTorch.Source: D. Simulation and experiment setup · p. 3
  • Mode A uses TinyLlama-1.1B and the Fed-Equilibrium cascade with λk derived from the Pareto knee at wus = 2.0.The mode is presented as the cooperative solution.Source: D. Simulation and experiment setup · p. 3
  • Mode B uses size-proportional FedAvg aggregation, while Mode C assigns equal influence to all 13 nodes.The setup characterizes Mode B as non-cooperative and Mode C as naïve egalitarianism.Source: D. Simulation and experiment setup · p. 3

III. RESULTS

The results section evaluates Fed-Equilibrium across semantic harmonization, convergence under imbalance, and resilience across three experimental modes.

III. RESULTS

  • The evaluation covers semantic harmonization, Pareto-optimal convergence under imbalance, and resilience across three modes.Source: §III · p. 4
  • One evaluation dimension is semantic harmonization.Source: §III · p. 4
  • The other dimensions are convergence under imbalance and resilience across the three modes.Source: §III · p. 4

A. Data harmonization and semantic alignment

The results report complete variable harmonization between the Canadian and U.S. datasets, then examine the Pareto frontier and its weighting trade-offs.

A. Data harmonization and semantic alignment

  • The OMOP mapping pipeline unified 100% of variables across 37,287 Canadian CNODES records and 974 U.S. SyntheticMass records despite differences in coding and clinical text.The Canadian data use standardized ATC codes, while the U.S. data contain unstructured clinical text.Source: A. Data harmonization and semantic… · p. 4

B. Empirical identification of the Pareto frontier

  • At w = 2.0, minority loss falls from 0.857 to 0.340 while majority loss remains comparable at 0.981 versus 0.985.The paper identifies this as the topological Pareto knee point and reports that it decouples optimization from hub volume dominance.Source: Figure 3 · p. 4, B. Empirical identification of the… · p. 4
  • Standard FedAvg at w ≈ 0.03 has majority loss 0.985 and minority loss 0.857; excessive upweighting at w > 10 raises majority loss above 1.6.The passage describes a convex frontier between majority and minority loss.Source: B. Empirical identification of the… · p. 4
  • The sensitivity analysis sweeps wus to identify where marginal fairness gains no longer justify marginal costs in global utility.Source: Figure 4 · p. 4

C. Synergistic resilience validation (TPC)

Fed-Equilibrium’s Pareto weighting improved minority US convergence while preserving majority CA performance, outperforming standard aggregation strategies. The evaluation compares learning trajectories and final loss distributions.

C. Synergistic resilience validation (TPC)

D. System stability and architectural synergy

The cooperative validation reports strong directional alignment across the network, followed by TPC reweighting that improves minority convergence without reducing majority performance.

D. System stability and architectural synergy

  • Across 13 nodes, mean cos θ was 0.99 ± 0.01, and the GQA gatekeeper accepted the cooperative updates as directionally consistent benign gradients.The authors report that the funnel established a unified trust region without rejecting valid clinical insights.Source: D. System stability and architectural… · p. 5
  • The two stages serve distinct roles: GQA checks update integrity, while TPC addresses the remaining imbalance between large Canadian hubs and US spokes.The geometric check alone does not prevent majority updates from overwhelming minority updates.Source: D. System stability and architectural… · p. 5
  • TPC amplified underrepresented US spokes within the aligned benign-gradient pool, yielding final loss 0.3390 for US nodes and 0.9926 for majority hubs.The reported outcome is presented as bridging Stage I acceptance of benign data and Stage II accommodation of underrepresented participants.Source: D. System stability and architectural… · p. 5

IV. DISCUSSION

Fed-Equilibrium combines semantic harmonization, geometric quality assurance, and topological Pareto control to address knowledge dominance left unresolved by prior geometric defenses. The discussion presents these layers as a route to secure and equitable clinical federated networks.

IV. DISCUSSION

  • Prior geometric defenses supplied a security baseline but left knowledge dominance unresolved; Fed-Equilibrium adds topological control to pursue fairness alongside security.The framework is described as operating through semantic harmonization, geometric quality assurance, and topological Pareto control.Source: §IV · p. 5

A. Strengths: A three-layer synergistic architecture for clinical fairness

  • At the Pareto knee (w = 2.0), minority loss drops from 0.857 to 0.340 without degrading the majority baseline.The discussion characterizes this as escaping FedAvg’s Pareto inefficiency; the knee was identified through a static parameter sweep.Source: A. Strengths: A three-layer… · p. 5
  • GQA establishes a stabilized manifold with mean cos θ = 0.99 ± 0.01 across 13 nodes before TPC rebalances contributions.The discussion says this sequence ensures only directionally consistent minority updates are amplified.Source: A. Strengths: A three-layer… · p. 5
  • The OMOP-CDM pipeline unified 100% of heterogeneous variables from Canadian ATC codes and U.S. free-text narratives into a shared instruction-tuning space.The discussion presents semantic harmonization as a prerequisite for downstream geometric and topological operations.Source: A. Strengths: A three-layer… · pp. 5–6
  • The static parameter sweep provides an empirical baseline, while dynamic weighting that recalibrates influence during real-time training remains future work.Source: A. Strengths: A three-layer… · p. 5

B. Limitations and future directions

The conclusion reports deep minority-node convergence without degrading majority utility. Future directions identify boundaries in the data setting, TPC recalibration, and adversarial validation.

B. Limitations and future directions

  • The framework currently operates on structured EHR and administrative claims data; multimodal, real-world deployment remains future work.Anticipated challenges include missing data, label noise, and site-specific coding, which will require adaptive TPC extensions.Source: B. Limitations and future directions · p. 6
  • TPC currently uses an offline sensitivity sweep of about 1.8 GPU-hours to set a fixed Pareto-alignment coefficient.Future work proposes online recalibration using client-graph topology and Bayesian hyperparameter search to reduce overhead.Source: B. Limitations and future directions · p. 6
  • Validation under cooperative conditions leaves robustness boundaries under adaptive Byzantine attacks to be stress-tested.Proposed tests include gradient collusion and targeted label-flipping.Source: B. Limitations and future directions · p. 6

V. CONCLUSION

  • Minority nodes reached deep convergence at loss 0.340 without degrading majority utility at loss 0.985.The authors conclude that global generalizability can coexist with local representation.Source: §V · p. 6
Loading 2609.11937v1…