PaperlayerView exact revision on arXiv

Source-linked AI summary

RegionFed: Federated Learning for Personalized Query Understanding in Heterogeneous Retail Environments

Quoc H. Nguyen, Ali Lafzi, Abhijeet Phatak, Siddharth Pratap Singh, Rohit Upadhyay, Yogananda Domlur Seetharama, Chittaranjan Tripathy

arXiv:2609.05403v1cs.LGcs.AI

TL;DR

Regional heterogeneity limits global federated models, while parameter-level personalization can collapse on transformers. RegionFed uses gradient conflict to adapt regionally and route personalization strategies, achieving near-centralized accuracy across architectures with differential privacy.

  • Problem

    Retail query data varies across regions, but standard FL sacrifices regional performance and parameter-level personalization becomes unstable on modern transformers.

  • Method

    RegionFed performs gradient-level regional personalization, using ℓ2 gradient conflict to diagnose heterogeneity, select strategies, and control adaptation strength.

  • Results

    92.27% overall accuracy is achieved across three datasets and four architectures, within 1σ of the centralized 92.04% regional-weighting upper bound by Δ=0.23pp, with ε≈0.60-DP.

  • Takeaways & Limitations

    Gradient-level operations remain stable where parameter-level corrections collapse, enabling deployment across transformers and CNNs without code changes.

  • Takeaways & Limitations

    RegionFed requires pre-defined regions, does not evaluate multimodal or state-space-model tasks, and occasionally misroutes with Dynamic strategy selection.

Abstract

from arXiv · show

Retail search systems serve diverse geographic regions with distinct query patterns, vocabularies, and product preferences, creating significant data heterogeneity that challenges both privacy-preserving training and model personalization. Federated learning offers a natural solution for privacy, but standard FL methods produce global models that sacrifice regional performance, while existing personalized FL approaches operate at the parameter level and catastrophically collapse on modern transformers (below 10\% accuracy on T5) due to tied embeddings and LayerNorm interactions. We introduce RegionFed, an \textit{architecture-robust} federated learning framework that sidesteps this failure by operating entirely at the gradient level. RegionFed uses the $\ell_2$ conflict between regional and global gradients as a unified signal that (i) diagnoses heterogeneity, (ii) routes each region to the cheapest sufficient personalization strategy, and (iii) adaptively controls personalization strength. Because it treats models as differentiable black boxes, RegionFed deploys on T5-Small, T5-3B, RoBERTa, and CNN with zero code changes, providing large gains on transformers (where parameter-level methods collapse) and consistent improvements on CNNs. Across three public datasets (Amazon ESCI, Amazon Reviews, LEAF-FEMNIST) and four architectures, RegionFed-Meta achieves 92.27\%, closing the gap to the privacy-violating centralized upper bound (Centralized + Regional Weighting: 92.04\%, $Δ$=0.23pp, within 1$σ$) while providing $(ε{\approx}0.60)$-differential privacy and $\mathcal{O}(1/\sqrt{T})$ convergence.

1 Introduction

RegionFed addresses regional heterogeneity and transformer instability in federated retail query understanding by personalizing at the gradient level. Its adaptive framework selects regional strategies and improves performance across architectures while preserving privacy.

  • Motivation: Regional shopping behaviors, vocabularies, and product preferences create heterogeneous query-understanding data that cannot simply be centralized under privacy and sovereignty constraints.Standard federated learning provides privacy-preserving distributed training, but existing clustered and personalized methods do not fully address heterogeneous retail environments.
  • Motivation: Standard FL produces one global model that sacrifices regional accuracy, while parameter-level personalization becomes unstable on transformers because of shared embeddings, attention coupling, and LayerNorm interactions.The motivating example is that “thongs” denotes different products in Australia and the US.
  • Core idea: Gradient-level personalization treats models as black-box function approximators, using regional gradient conflicts to provide architecture-robust adaptation at the regional level.Regions pool enough clients for stable gradient estimates, while users within regions share systematic patterns.
  • Core idea: RegionFed combines gradient conflict analysis with hierarchical regional structure to diagnose heterogeneity, select personalization strategies, and control adaptation strength under differential privacy.The framework uses ℓ2 gradient conflict as its sole adaptive signal rather than directly manipulating model parameters.
  • Reported results: 12–25pp improvement over FedAvg/FedProx is reported on T5-Small, while all four tested parameter-level baselines collapse below 10% on transformers.The transformer-specific failure is additionally confirmed on LEAF-FEMNIST, where SCAFFOLD reaches 79.52% on CNN.
  • Reported results: RegionFed is evaluated across RoBERTa-Base, T5-3B, and CNN, with cross-domain validation on Amazon Reviews and sustained advantage over 50 rounds.Results are averaged over 5 seeds.

2 Challenges and Problem Formulation

The paper formulates retail federated learning as a two-level global–regional optimization problem under heterogeneous client data. Regional structure captures systematic semantic differences while preserving private client data and task-level objectives.

  • Challenges: Regional semantic differences, product availability, and seasonal timing cause client drift and an irreducible heterogeneity error for a single global model.The paper motivates region-aware personalization because one global model inevitably sacrifices regional accuracy.
  • Problem formulation: The system contains N private clients distributed across M regions, with each client u in region r holding private data Dᵣᵤ.Regions contain nᵣ clients and N = Σᵣ nᵣ.
  • Problem formulation: RegionFed optimizes a two-level global-plus-regional objective over T communication rounds, while user-level parameters are computed through on-device local adaptation.This separates federated regional optimization from post-deployment user personalization.
  • Problem formulation: The objective weights regions by pᵣ = 1/M and combines K task losses through task weights λₖ, with regional data size defined over participating clients.The setup uses equal-sized region weights.
  • Problem formulation: Task-level losses are explicitly connected to the regional and user losses used by later personalization mechanisms.This definition bridges the multi-task query-understanding objective with subsequent regional and user-level training terms.

3 RegionFed: System Architecture

RegionFed uses a three-layer global–regional–client hierarchy in which DP-noised gradients drive regional adaptation, strategy selection, and personalization weights. The same gradient-conflict signal controls both which strategy a region uses and how strongly it adapts.

  • Architecture: The architecture has a global server, regional coordinators, and local clients, with coordinators maintaining regional models and optimizing adaptive weights and ρ.Clients train on private data and communicate DP-noised gradient updates.
  • Architecture: User-specific models are composed at deployment from global, regional, and locally fine-tuned components, with local adaptation remaining on-device and outside federated communication rounds.The composition is θᵤ = θ + αᵣ · θᵣ + αᵤ · θlocalᵤ.
  • Privacy-preserving gradients: Regional gradients are securely aggregated, clipped, and Gaussian-noised before global aggregation, yielding (ε, δ)-differential privacy with ε ≈ 0.60 for T=50, q=0.1, and δ=10^-5.The clipping threshold is C = 1.0 and σdp = 4.0.
  • Adaptive personalization: Regional personalization combines global knowledge with a region-specific adaptation weighted by αᵣ ∈ [αmin, 1.0], where αmin = 0.5 by default.The lower bound prevents regional adaptation from being discarded entirely.
  • Personalization strategies: RegionFed supports Grad, Interp, Meta, and Dynamic strategies, with per-region adaptation intensity ρ optimized by golden section search.All regions converge to similar values, with mean ρ = 0.0344.
  • Adaptive strategy selection: Dynamic routing selects the cheapest sufficient strategy using gradient conflict, heterogeneity, and regional dataset size.The selection compares dᵣ = ∥ĝᵣ − ĝ∥₂ and hᵣ = KL(Pᵣ∥Pglobal) against configured thresholds.
  • Training procedure: Each round broadcasts the global model, selects regional strategies, computes DP-noised gradients, and updates regional adaptations; post-processing preserves the privacy guarantee.The gradient conflict ∥ĝᵗᵣ − ĝᵗ∥₂ drives both strategy choice and αᵣ.

4 Experimental Setup

The experiments evaluate RegionFed on retail NLP and FEMNIST vision tasks using multiple architectures, datasets, baselines, and regional metrics. The setup emphasizes cross-architecture robustness and comparison with centralized and federated alternatives.

  • Datasets: The benchmark includes Amazon ESCI, Amazon Reviews, and LEAF-FEMNIST, spanning retail search, sentiment classification, and writer-partitioned character recognition.The datasets cover NLP and vision under regional or client heterogeneity.
  • Tasks and metrics: Query understanding uses intent classification, spell correction, and named entity recognition in a unified text-to-text framework.Overall Accuracy is the unweighted mean of intent accuracy, spell-correction accuracy, and NER token-level F1.
  • Tasks and metrics: Regional Robustness Score measures worst-region uplift, with higher RRS and lower inter-region standard deviation indicating more equitable performance.The metric averages regional accuracies across M regions.
  • Architectures: Architectures include T5-Small, T5-3B, RoBERTa-Base, and a 2-convolution, 2-fully connected CNN.T5-Small has 60.5M parameters, while T5-3B and RoBERTa-Base test larger transformer settings.
  • Baselines: Baselines cover centralized upper bounds, FedAvg, FedProx, region-aware aggregation, parameter-level personalization, layer splitting, transformer-specific methods, local fine-tuning, and RegionFed variants.The comparison isolates hierarchical structure from gradient-conflict adaptation through Region-Aware FedAvg.
  • Baseline results: 83.21 overall is reported for Region-Aware FedAvg, compared with 80.04 for FedAvg and 67.13 for FedProx.The values are reported as overall accuracy in the standard FL comparison.
  • Baseline results: 91.54% is achieved by FedTP on one reported comparison, while Local Fine-tuning records 79.53 overall and FedBABU records 84.12 overall in the table.These values are table-level overall accuracy results for the corresponding baselines.

FedAvg SCAFFOLD FedBABU FedTP RF-Meta

Across architectures, RegionFed-Meta performs strongly on transformers and CNNs, while parameter-level methods collapse on transformers but remain viable on CNNs.

  • Transformer results: 94.12% on ESCI T5-3B and 91.62% on ESCI RoBERTa-Base show strong RegionFed-Meta performance on transformers.The corresponding centralized upper bounds are 94.81% and 91.38%, respectively.
  • Parameter-level methods: SCAFFOLD reaches 79.52% on FEMNIST CNN but collapses on transformer settings, confirming that its failure is architecture-specific.The paper states that SCAFFOLD’s transformer collapse is below 10%, while its CNN result remains substantially higher.
  • Architecture generality: RegionFed uses the same framework across T5-Small, T5-3B, RoBERTa, and CNN evaluations, whereas FedTP requires a transformer-specific hypernetwork and is not applicable to CNNs.The training configuration also reports that RegionFed at η=10−4 achieves 91.14%, indicating its gains are not solely attributed to the selected learning rate.

5 Results and Analysis

RegionFed combines gradient-conflict adaptation with regional personalization to improve accuracy, transformer robustness, heterogeneity resilience, query-level predictions, and convergence under differential privacy.

  • Overall performance: 92.27% is RegionFed-Meta’s overall accuracy, within 0.23pp of the 92.04% centralized regional-weighting upper bound.The reported comparison places RegionFed-Meta within statistical noise of the stronger centralized upper bound.
  • Standard FL: FedAvg reaches 80.18% and FedProx 67.18%, remaining 12–25pp below centralized training.FedProx’s uniform parameter regularization is reported to prevent the specialization needed for regional adaptation.
  • Transformer robustness: SCAFFOLD, pFedMe, Ditto, and APFL all fall below 10% on T5-Small across E∈{1, 5, 10, 40}.Ablations attribute the collapse to interactions involving T5’s shared embeddings and LayerNorm; gradient-level operations avoid this parameter-level failure.
  • Heterogeneity robustness: RegionFed-Meta degrades 6.5pp under increased heterogeneity, compared with FedAvg’s 21.6pp degradation.The paper reports 92.0%→85.5% for RegionFed-Meta and describes this as 3.3× less degradation.
  • Convergence: RegionFed plateaus above 91% by round 5, whereas FedAvg reaches approximately 82% by round 20.The convergence figure states that the persistent gap cannot be closed by simply adding rounds.
  • Theory and privacy: Theorem 5.1 provides O(1/√T) convergence, while the privacy analysis reports ε≈0.60 under the stated differential-privacy parameters.The paper also reports RegionFed-Meta retaining above 91% accuracy at ε=0.30 and achieving 88.15% at ε=0.15.

6 Conclusion and Limitations

RegionFed uses gradient-level personalization to remain stable across transformers and CNNs, achieving strong accuracy while preserving differential privacy. Its limitations include reliance on predefined regions, unevaluated model and task settings, and occasional dynamic-routing errors.

  • Conclusion: RegionFed personalizes through gradient conflict rather than direct parameter manipulation, enabling deployment across transformers and CNNs without code changes.Its gradient-level operations remain stable where parameter-level corrections collapse.
  • Empirical scope: The evaluation spans three public datasets and four architectures, with task-level gains for NER and intent but a centralized advantage on spell correction.NER reaches 99–100% versus centralized 97.2%, intent reaches 83.8% versus 77.2%, and spell reaches 92.0% versus 100%.
  • Limitations: The method requires predefined regions, leaves multimodal tasks and state-space models unevaluated, and can occasionally misroute dynamic strategy selection.Region misspecification incurs a reported 4.2pp loss.
  • Contribution: RegionFed’s gradient-level approach combines architecture robustness with regional hierarchical aggregation, unlike prior methods that remain architecture-specific or parameter-level.The framework is presented as filling a gap in architecture-agnostic federated learning for retail search with transformers.
  • Empirical scope: RegionFed improves transformer and CNN performance across tasks, including 68.94% versus centralized 70.21% and 85.21% on CNNs.Parameter-level methods collapse below 10% on the transformer task, whereas RegionFed remains functional.
  • Conclusion: 92.27% overall accuracy slightly exceeds centralized training at 91.72%, with a modest 0.55pp difference.The reported result is interpreted as eliminating the traditional federated-learning performance penalty.

C.2 Finding 2: Why Standard FL Methods Are Insufficient

Standard global and parameter-level federated methods underperform on heterogeneous retail queries, especially on transformers. RegionFed addresses this through gradient-level regional adaptation, with regional aggregation improving robustness and strategy selection balancing accuracy against computational cost.

  • Gradient conflict: Negative gradient similarity causes averaging to cancel useful information, producing compromise updates that may be suboptimal for every region.RegionFed uses L2 distance because it captures both gradient direction and magnitude differences.
  • Standard FL limitations: Uniform FedProx regularization prevents fast-drifting attention parameters from specializing while offering little benefit for slowly changing embeddings.The mismatch arises because one regularization strength is applied regardless of functional role.
  • Transformer failure: Parameter-level personalization collapses below 10% on T5-Small, whereas tied-embedding and LayerNorm ablations identify the main instability sources.Combining untied embeddings and frozen LayerNorm recovers SCAFFOLD to 71.83%.
  • Regional adaptation: Gradient-level local fine-tuning reaches 88.84%, but regional aggregation raises regional robustness from 79.53% to 91.62%.Regional aggregation uses approximately 2,000 samples per region versus approximately 200 per client.
  • Strategy trade-offs: RegionFed strategies achieve 91–92% overall accuracy, with Grad minimizing overhead, Meta maximizing accuracy, and Dynamic routing regions among strategies.The production guidance recommends Grad under compute constraints, Meta when accuracy is critical, and Dynamic with a Grad fallback.
  • Strategy limitations: Dynamic routing misfires on Grocery at 78.2% because of strategy-selection hysteresis, motivating fallback after two rounds without validation improvement.Grad wins 4/8 regions, Meta 3/8, and Dynamic 1/8.
  • Regional performance: Grad and Meta remain consistent across regions at 90.28%–93.40%, while Interp reaches only 83.56%–87.21%.Grad is optimal in 4/8 regions, Meta in 3/8, and Dynamic in Sports.

D.2 Adaptation Intensity Analysis

RegionFed’s adaptation strategies trade accuracy, consistency, computational cost, and tuning complexity. Gradient-based strategies provide efficient default behavior, while Meta and Dynamic expose distinct benefits and deployment risks across heterogeneous tasks and regions.

  • Adaptation intensity: Grad and Meta use constant ρ=0.0344 across regions, whereas Interp varies from 0.10 to 0.34 and requires more region-specific tuning.Fashion and Home require the largest Interp intensity, while Beauty and Electronics require the smallest.
  • Strategy trade-offs: 91.92%–92.27% overall accuracy spans all four strategies, but computational requirements, memory footprint, and deployment complexity differ substantially.Grad is the lowest-cost strategy, while Meta incurs additional wall-clock cost from its MAML inner loop.
  • Strategy trade-offs: Grad achieves 91.92% overall accuracy with 91.52% RRS, winning 4/8 regions and providing the recommended default for constrained resources and moderate heterogeneity.Its single gradient computation minimizes overhead while maintaining consistent regional performance.
  • Strategy trade-offs: Meta achieves the highest overall accuracy at 92.27% and best regional consistency at 91.62% RRS, making it preferable when maximum specialization is critical.Meta wins 3/8 regions with complex query patterns and requires 1.15× FedAvg wall-clock time.
  • Regional performance: All eight case studies show 5.0%–7.8% personalization gains over the global baseline, with Meta strongest in complex high-variance categories and Grad competitive at lower cost.Winning margins are small, ranging from 0.0% to 0.4%, while Grad wins 4/8 cases.
  • Dynamic selection: Dynamic strategy reaches 78.2% on Grocery intent, below Grad at 91.1%, because declining conflict triggers interpolation that damages Grocery-specific patterns in a feedback loop.The resulting misselection motivates monitoring infrastructure for production deployment.

F.5 Controlled Ablation: Isolating the Collapse Cause

Controlled ablations show that parameter-level personalization collapses on T5 because tied embeddings and LayerNorm interact badly with full-parameter corrections. RegionFed avoids these architecture-specific failure modes through gradient-level updates, while layer splitting remains stable but less accurate and less robust to heterogeneity.

  • Controlled ablation: Untying embeddings raises SCAFFOLD from 8.73% to 42.61%, while freezing LayerNorm raises it to 38.47%, identifying both as primary instability sources.The controlled variants modify one architectural feature at a time under otherwise identical settings.
  • Controlled ablation: Combining untied embeddings with frozen LayerNorm recovers SCAFFOLD to 71.83%, with the remaining gap attributed to attention coupling.LoRA-only updates reach 76.24%, further linking collapse to full-parameter updates on coupled transformer components.
  • Mechanistic interpretation: RegionFed’s gradient-difference signal avoids interactions with parameter tying, normalization layers, and parameter structure that destabilize control-variate methods.The signal is computed as a single vector operation rather than a parameter-structured correction.
  • Layer-splitting comparison: FedBABU avoids catastrophic collapse and reaches 84.53% on T5-Small, but remains below RegionFed-Meta at 92.27% because head-only personalization cannot adapt the transformer body.Regional semantic differences are encoded in attention patterns and intermediate representations, not only the output head.
  • Layer-splitting comparison: RegionFed-Meta degrades 6.5pp under high heterogeneity versus FedBABU’s 17.4pp degradation, showing stronger robustness to distribution shifts affecting intermediate representations.FedBABU remains architecture-stable across transformers and CNNs but does not match RegionFed’s cross-architecture accuracy.

H Component Ablation Study

The ablation study identifies adaptive personalization weights as the most important component, while hierarchical personalization and golden-section regularization add complementary gains. Sensitivity analyses support calibrated defaults for temperature, centering, minimum personalization, and DP clipping.

  • Component Contributions: Adaptive α contributes +1.12%, or 39.7% of total improvement, making it the most critical component.It resolves gradient conflicts through dynamic balancing.
  • Component Contributions: Hierarchical Personalization contributes +0.93%, while combining it with adaptive α yields +2.29% through a 1.29× synergy.The two-level optimization captures regional patterns and client preferences.
  • Synergistic Effects: The three components exceed their individual sum by +0.32%, representing 12.8% synergy through interacting interpolation, balancing, and regularization effects.The reported feedback loop links hierarchy, ρ optimization, and α balancing.
  • Sensitivity Analysis: Temperature sensitivity is robust across the tested range, with accuracy varying by only ±1.2% and τ = 1.0 providing the best sensitivity-stability balance.Lower temperatures can overcorrect, while higher temperatures dampen conflict signals.
  • DP Clipping Sensitivity: C=1.0 is optimal among DP settings, achieving 92.27%, only 0.26pp below the non-private upper bound of 92.53%.All finite clipping settings retain ϵ=0.60; clipping changes utility rather than privacy under fixed noise-to-sensitivity ratio.

I.1 Convergence Analysis

RegionFed’s theory establishes convergence for both global and personalized models under heterogeneous data, with the rate depending on smoothness, variance, and regional dissimilarity. Its privacy analysis gives a formal differential-privacy guarantee under secure aggregation and user-level sampling.

  • Assumptions: The convergence analysis assumes smoothness, bounded stochastic-gradient variance, bounded regional gradient dissimilarity, and bounded region-specific adaptation.These assumptions define the conditions under which the theorem applies.
  • Assumption Validation: The measured regional adaptation norm averages 0.034 of the global norm and reaches at most 0.052, supporting the bounded-adaptation assumption.The reported validation uses all 8 regions across 50 communication rounds.
  • Convergence Rate: RegionFed achieves an O(1/√T) convergence rate for both global and personalized models under heterogeneous data.The result matches the optimal rate stated for non-convex federated optimization.
  • Personalization Bound: The personalization bound links the regional loss to global-regional dissimilarity and bounded adaptation, providing a theoretical basis for adaptive αr.The analysis states that the optimal weight is primarily governed by regional dissimilarity when the adaptation bound is small.
  • Privacy Guarantee: With C=1.0, σdp=4.0, T=50, q=0.1, and δ=10^-5, the moments accountant yields ϵ ≈0.60.The guarantee uses secure aggregation and user-level sampling, with DP noise applied at the regional level.

J Privacy-Utility Tradeoff Analysis

The privacy-utility analysis evaluates RegionFed-Meta across increasing privacy strength and relates its robustness to regional aggregation. The section also frames topology-based guidance as an optional mechanism for preserving global structure during adaptation.

  • Privacy-Utility Frontier: RegionFed-Meta maintains >91% accuracy from ϵ=2.40 down to ϵ=0.30, degrading only 0.79pp across this 8× privacy range.The comparison uses T5-Small on Amazon ESCI with fixed training and privacy settings.
  • Privacy-Utility Frontier: At ϵ=0.15, RegionFed-Meta achieves 88.15%, while its advantage over FedAvg increases to +13.94pp under stronger privacy.The reported advantage rises from +11.69pp at ϵ=2.40.
  • Noise Robustness: Regional aggregation averages DP noise across clients within each region, providing an implicit noise-reduction effect proportional to regional client count.This is presented as the explanation for RegionFed’s greater noise tolerance.
  • Topology-Guided Adaptation: Topology loss compares global and regional neighborhood graphs to balance regional specialization against preservation of global structural knowledge.Its weight controls the trade-off between regional task performance and structural similarity.

K.1.4 Experimental Analysis

Additional experiments show that RegionFed transfers across architectures and model scales, while topology guidance and asynchronous operation introduce practical trade-offs. The section also documents assumptions and scope boundaries for deployment and future extensions.

  • Computational Trade-offs: Topology loss costs O(N^2d) per evaluation and adds 50ms for N = 200 and d = 512, making it suitable for offline optimization but potentially prohibitive for real-time deployment.The default adaptive-weight approach avoids this overhead while achieving 98% of the performance.
  • Optimization Assumption: The golden-section search relies on a unimodal regional-loss assumption that was empirically verified across all 8 regions and multiple rounds.Grid-search optima matched golden-section results within ϵ = 0.02.
  • Cross-Architecture Validation: RegionFed-Grad improves RoBERTa-Base accuracy from 81.2% with FedAvg to 91.5%, a 10.3pp gain without architecture-specific adapters.The evaluation uses an encoder-only architecture for intent classification.
  • Cross-Scale Validation: RegionFed-Meta achieves 94.12% on T5-3B, a +1.85pp improvement over the T5-Small result of 92.27%.The personalization overhead remains constant relative to base training because it uses scalar projections of clipped gradient norms.
  • Architecture Robustness: The same gradient-level algorithm generalizes across encoder-only and encoder-decoder models and from 60.5M to 3B parameters.The experiments use identical personalization hyperparameters across these architectures and scales.
  • Deployment Scope: Full asynchronous evaluation under real-world network conditions remains future work, despite proposed staleness bounds and production mechanisms.The current implementation uses synchronous communication rounds.

L.3 Proof of Theorem I.7 (Privacy Guarantee)

The privacy guarantee analyzes RegionFed with moments accounting and privacy amplification across user and regional sampling. In the practical deployment, regional noise and secure aggregation yield ε≈0.60 while adaptive personalization does not consume additional privacy budget.

  • Privacy analysis: RegionFed analyzes differential privacy through moments accounting and amplification by sampling at user and regional levels.The general proof considers two-level composition, while the practical algorithm applies noise at the regional level and uses secure aggregation for user-to-region privacy.
  • Privacy amplification: The effective end-to-end sampling ratio is q · p because a user must be selected within a selected region.The proof explicitly combines region selection probability p with within-region user sampling probability q.
  • Adaptive personalization: Adaptive personalization weights preserve the privacy guarantee because they are computed from noised gradients through post-processing.The post-processing property prevents additional privacy loss when the weights use differentially private outputs without accessing private data.
  • Practical guarantee: ε≈0.60<1 is achieved for T=50, q=0.1, σdp=4, and δ=10^-5 while maintaining high utility.The paper identifies this as the main evaluation setting and uses it to complete the privacy guarantee.
  • Practical guarantee: ε=0.58 is obtained with an RDP accountant at optimal order α*≈17, consistent with the analytic bound.The numerical verification supports the practical privacy calculation under the stated deployment and accounting choices.
  • Limitations and ethics: The scope remains bounded by predefined regions, limited validation domains, occasional Dynamic-strategy failures, and the need for further privacy and statistical analysis.The authors also identify potential regional bias amplification and discriminatory pricing risks without fairness auditing.
Loading 2609.05403v1…