Source-linked AI summary
vFedProtoQNAS: Prototype-Guided Personalized Quantum Neural Architecture Search for Virtual Federated Learning
Seok Bin Son, Samuel Yen-Chi Chen, Soohyun Park, Joongheon Kim
TL;DR
Device heterogeneity makes a single shared QNN architecture unsuitable, and personalized QNAS leaves parameters structurally misaligned for averaging. vFedProtoQNAS instead shares class-wise prototypes and uses global prototypes to refine client representations; it reports higher accuracy than FedAvg and class-consistent alignment.
Problem
Device constraints make identical QNN architectures unsuitable across clients, while personalized architectures leave parameter indices structurally misaligned for averaging.
Source: §1 · p. 2, §3.1 · p. 4Method
Each client searches for and trains a personalized QNN, shares class-wise prototypes for server aggregation, then refines local representations using global prototypes without aggregating model parameters.
Source: §1 · p. 2, §3.5 · p. 5Results
Accuracy is 91.25 ± 0.74% for vFedProtoQNAS versus 87.55 ± 2.38% for FedAvg, alongside reported class-consistent representation alignment.
Source: §1 · p. 2, §4.2 · p. 7Takeaways & Limitations
Prototype-based representation sharing supports collaboration across client-specific QNN architectures without requiring structurally aligned models.
Source: §5 · p. 8Despite different internal architectures, clients must use the same number and ordering of measured observables for prototype computation.
Source: §3.1 · p. 3
Abstract
from arXiv · showhide
Quantum federated learning (QFL) has emerged as a promising approach for collaboratively training compact quantum neural networks (QNNs) over distributed private data on resource-constrained devices. However, differences in device capabilities make a single shared QNN architecture unsuitable for all clients. While personalized quantum neural architecture search (QNAS) allows each client to select a device-specific QNN, averaging parameters across structurally different QNN architectures mixes semantically inconsistent circuit operations. To address this, prototype-guided personalized QNAS for virtual FL (vFedProtoQNAS) is proposed, where model parameters are never aggregated across clients and federated collaboration is achieved through class-wise prototype sharing. Each client independently searches and trains a client-specific QNN, computes class-wise local prototypes from latent representations, and refines them using global prototypes from the server as federated semantic anchors. Experiments demonstrate that vFedProtoQNAS improves accuracy by 3.70\% over FedAvg and enhances class-consistent representation alignment.
1 Introduction
Heterogeneous client resources make a single shared QNN architecture unsuitable, while personalized QNAS creates parameter-alignment problems for aggregation. vFedProtoQNAS enables client-specific search and training while collaborating through class-wise prototypes and global-prototype refinement.
1 Introduction
- A single architecture may exceed low-resource clients’ budgets or limit resource-rich clients’ learning, motivating adaptable designs for heterogeneous FL.Clients differ in computational resources, memory capacity, and energy budgets.Source: §1 · pp. 1–2
- Personalized QNAS lets clients select QNNs for device constraints and local data, but parameter averaging can mix different operations at matching indices.Circuit depth and gate type affect QNN performance and efficiency.Source: §1 · p. 2
- vFedProtoQNAS replaces parameter aggregation with class-wise prototype sharing and uses server-generated global prototypes as semantic anchors for local representations.Clients compute class-wise local prototypes from latent features; the server aggregates them using sample weights and returns the global prototypes for refinement.Source: §1 · p. 2
- Class-wise prototype sharing and global-prototype refinement induce class-consistent representation alignment and improve classification over major baselines.Clients retain their personalized architectures and local parameters.Source: §1 · p. 2
2 Preliminary
The preliminary section describes QNN components and quantum architecture search, then explains how conventional QFL aggregates client parameters under a shared-architecture assumption.
2 Preliminary
- A QNN encodes classical inputs, processes them with a trainable parameterized quantum circuit, and measures observables for classical readout.The circuit uses one-qubit rotations and two-qubit gates for entanglement; parameters are updated using the parameter-shift gradient rule.Source: §2 · p. 3
- QNAS searches circuit designs that trade representation capacity against cost, noise sensitivity, and trainability, with one-shot NAS using a shared supernet.The search spans choices including gates, placement, depth, and entanglement topology.Source: §2 · p. 3
- Conventional QFL averages uploaded client QNN parameters, an operation that assumes shared architectures and aligned parameter spaces.With different architectures, a parameter index may denote different gates or circuit paths, making aggregation inconsistent.Source: §2 · p. 3
3 vFedProtoQNAS Framework
vFedProtoQNAS searches for client-specific QNNs and replaces parameter averaging with class-wise prototype sharing in a common latent interface. Global prototypes guide local refinement while each client retains its architecture and parameters.
3.1 Problem Formulation
- Different client architectures can make parameter averaging structurally inconsistent, so vFedProtoQNAS instead collaborates through class-wise prototypes in a common latent interface.The interface uses the same number and ordering of measured observables across clients.Source: §3.1 · pp. 3–4
3.2 Architecture
- Clients search and train personalized QNNs, share class-wise prototypes and sample counts, then refine local parameters using server prototypes while keeping architectures fixed.Raw data and model parameters remain local.Source: §3.2 · p. 4
3.3 Client-Specific ProxylessQNAS
- Each client selects an executable QNN within its device-feasible search space by balancing validation loss against architecture resource cost.Architecture logits guide cost-aware search, with λ_cost balancing predictive performance and resource efficiency.Source: §3.3 · p. 4
3.4 Prototype Computation and Aggregation
- Clients compute class-wise latent prototypes from local representations, and the server aggregates prototypes for each represented class using sample counts as weights.Clients transmit prototypes only for observed classes, along with their counts; raw data and model parameters are not transmitted.Source: §3.4 · pp. 4–5
3.5 Prototype-Guided Local Refinement
- Clients refine local parameters with a contrastive prototype-alignment loss that pulls representations toward the correct global class prototype and separates them from incorrect ones.The local classification loss is retained, while each personalized architecture remains fixed during refinement.Source: §3.5 · p. 5
4 Performance Evaluation
The evaluation tests vFedProtoQNAS on non-IID MNIST (0–3) across five clients, with one prototype aggregation and refinement step. It reports stronger baseline performance and improved representation alignment, while examining sensitivity to prototype weight and data heterogeneity.
4.1 Simulation Setup
- The MNIST (0–3) experiment uses five non-IID clients and one prototype aggregation followed by a 100-epoch refinement phase.Experiments were repeated over five random seeds; results are reported as mean ± standard deviation.Source: §4.1 · p. 5
4.2 Experimental Results
- Mean prototype margin improves from −0.143 to −0.007, while the positive-margin ratio rises from 43.61% to 53.51%.The margin distribution shifts toward positive values after refinement, indicating improved class-consistent alignment.Source: §4.2 · pp. 5–6, Figure 3 · p. 6
- Accuracy is 91.25 ± 0.53% at βproto = 0.05 and 91.25 ± 0.74% at βproto = 0.10, but falls to 90.94 ± 0.38% at βproto = 0.50.The paper attributes the decline at the larger weight to overly constraining client-specific representations.Source: Table 1 · p. 6, §4.2 · p. 6
- Accuracy is 91.15±0.76% at α = 1.0 and 91.16±0.93% at α = 10.0, versus 84.32 ± 11.77% at α = 0.5.At the stronger heterogeneity setting, some clients may have too few samples for certain classes, reducing local prototype reliability.Source: §4.2 · p. 6, Table 2 · p. 7
- Under the default βproto = 0.10 setting, vFedProtoQNAS achieves the highest performance among compared methods, with 91.25 ± 0.74% accuracy versus FedAvg’s 87.55 ± 2.38%, reported as mean±std over five seeds.Source: Table 3 · p. 7, §4.2 · p. 7
5 Conclusion
vFedProtoQNAS combines client-specific QNN architecture search with class-wise prototype sharing rather than parameter aggregation. The conclusion reports improved performance and class-consistent alignment, with architecture search and prototype-guided collaboration providing complementary benefits.
5 Conclusion
- vFedProtoQNAS lets each client search for and train a personalized QNN, replacing parameter aggregation with class-wise prototype sharing.Aggregated global prototypes refine client-specific representations while preserving personalized architectures and local specialization.Source: §5 · p. 7
- Personalized QNAS and prototype-guided collaboration provide complementary benefits: architectural adaptation and class-level semantic knowledge transfer.The conclusion also reports that prototype guidance improves class-consistent alignment by drawing representations toward correct-class prototypes and distinguishing them from incorrect ones.Source: §5 · p. 7
- Prototype-based representation-level collaboration supports heterogeneous QFL without requiring structurally aligned client models.The conclusion frames this as reconciling client-specific architectural personalization with federated collaboration.Source: §5 · p. 8