acceptodds
Under review as a conference paper at ICLR 2027

One-Shot Split Federated Learning with Prototype-Guided Synthetic Supervision

Abstract

Split Federated Learning (SFL) enables collaborative training across resource-constrained clients by partitioning models between clients and the server. However, under strict communication constraints, reducing client-server interaction to a single round severely limits available supervision, making effective training challenging under heterogeneous and imbalanced data distributions. In this work, we investigate one-shot SFL and identify the lack of reliable global supervision from limited intermediate representations as the key challenge. Existing one-shot federated learning methods typically assume access to complete local models or full model aggregation, making them difficult to apply to split federated learning where clients only optimize partial models. To address this issue, we propose Heterogeneous One-shot Split Federated Learning (HOSFL), a framework that constructs class-aware supervision from one-shot uploaded representations. HOSFL aggregates cross-client class prototypes and employs prototype-guided feature alignment and classifier-guided class consistency constraints to generate class-balanced synthetic samples, enabling effective model optimization within a single communication round. Extensive experiments on multiple benchmarks under diverse non-IID settings demonstrate that HOSFL achieves consistent performance improvements while significantly reducing communication overhead, with an average accuracy gain of 4.72% over strong SFL baselines. The anonymous code is available at https://anonymous.4open.science/r/HOSFL-10B3.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Split Federated Learning, One-Shot, Class Imbalance, Heterogeneous

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Y2hnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes HOSFL, a heterogeneous one-shot split federated learning framework designed to address communication constraints and class imbalance under non-IID client data distributions.

Experiments on CIFAR-10, CIFAR-100, and SVHN demonstrate improvements over several conventional split federated learning baselines under heterogeneous data distributions. The authors additionally provide ablation studies, runtime comparisons, client resource measurements, and empirical privacy evaluations.

The central contribution is an attempt to replace iterative split-learning communication with server-side synthetic supervision constructed from one-shot client uploads.

Critical strengths

S1. The paper addresses a practically relevant and challenging problem. Reducing client-server interactions in split federated learning is important for resource-constrained distributed systems. The combination of one-shot communication and severe client label heterogeneity presents a meaningful research problem. The proposed framework provides an end-to-end solution that integrates local representation learning, server-side classification, prototype-guided generation, and synthetic-data-assisted client refinement. This is a potentially useful direction for communication-efficient distributed learning.

S2. The reported empirical results are promising across several heterogeneous settings. The experiments cover three classification benchmarks, multiple Dirichlet heterogeneity levels, and different numbers of participating clients. The proposed method consistently reports higher classification accuracy than the evaluated baselines. The improvements are particularly notable when conventional SFL methods are directly restricted to one communication round. The additional client resource analysis and privacy experiments also reflect an effort to evaluate practical deployment considerations beyond predictive accuracy. These results suggest that synthetic supervision may be beneficial for addressing missing-class and class-imbalance problems in split learning.

Critical weaknesses

W1. The experimental protocol and baseline comparisons do not sufficiently establish the effectiveness of the proposed one-shot framework.

A major concern is the dependence on globally recoverable class labels. Section 3.1 and Appendix A state that clients upload intermediate representations together with global sample indices, allowing the servers to recover the corresponding labels.

However, it is unclear how the servers obtain and maintain this global sample-index-to-label mapping in a realistic federated learning scenario, where client data and associated labels are typically decentralized.

This assumption appears essential for training the supervised server-side classifier and constructing class-wise prototypes. Consequently, the paper needs to clarify whether HOSFL relies on additional centralized label information unavailable to competing methods.

Moreover, the comparisons with one-shot SFL, MultiSFL, and FedFSA are not fully convincing. In Table 1, the one-shot baselines achieve approximately 10–20% accuracy on ten-class datasets and 1–2% on CIFAR-100, approaching random classification performance.

These extremely weak results suggest that conventional iterative algorithms may not be suitable for direct conversion into one-shot protocols. Such comparisons do not necessarily establish superiority over methods specifically designed for one-shot learning.

HOSFL additionally benefits from extensive local self-supervised training, server-side optimization over uploaded sample-level representations, and a second client training stage using synthetic supervision. The comparison should therefore control the available supervision, training budget, and communication resources.

Importantly, reducing communication rounds does not automatically imply lower total communication volume. HOSFL uploads representations for local samples, transmits client model parameters, and distributes synthetic datasets back to clients. The paper does not provide sufficient quantitative evidence regarding the total transmitted bytes.

Without fair one-shot baselines and a complete communication-cost comparison, the main performance and efficiency claims remain insufficiently supported.

W2. The proposed feature aggregation and synthetic supervision mechanisms lack sufficient justification and targeted validation.

HOSFL independently trains client-side encoders using self-supervised objectives and then aggregates their intermediate representations to construct global class prototypes.

However, independently optimized client encoders may produce substantially different feature spaces, particularly under severe non-IID distributions.

Simply averaging representations from these encoders does not guarantee semantically meaningful global prototypes. Similarly, averaging client encoder parameters to construct the reference encoder does not necessarily produce representations compatible with the uploaded features.

Although Appendix B.2 provides a bound on prototype inconsistency, the result depends on client parameter divergence and does not establish that this divergence is sufficiently small in the experimental settings. The paper should empirically measure cross-client feature alignment and its effect on prototype quality.

There is also a concern regarding the prototype perturbation strategy in Eq. (11). Appendix B.3 shows that the expected squared feature-matching loss with stochastic prototype perturbation differs from the unperturbed objective only by a constant independent of the generated features.

Therefore, the perturbation does not introduce an explicit diversity-promoting objective in expectation. The claim that it preserves intra-class diversity or alleviates mode collapse requires stronger empirical evidence.

Furthermore, the generator is optimized using feature alignment and classifier consistency without an explicit constraint ensuring that generated samples follow the natural image distribution. This may encourage synthetic samples that satisfy the frozen classifier without preserving meaningful visual or semantic structure.

The current ablation study removes only broad components, rather than individually validating the prototype alignment, classifier guidance, perturbation mechanism, and second-stage refinement.

Consequently, it remains unclear which components are responsible for the reported improvements and whether the proposed synthetic supervision is robust.

Questions for authors

Q1. How are the global sample labels obtained by the servers? Q2. How is fairness ensured in the one-shot comparisons? The one-shot baselines in Table 1 achieve near-random accuracy, raising concerns about whether their original optimization mechanisms remain functional under the imposed protocol. Q3. How does HOSFL ensure compatibility among client feature spaces? Q4. What is the individual contribution of each generative supervision component? Q5. Could the authors clarify the reported numerical inconsistencies? For CIFAR-100 with α = 0.3, the full HOSFL model reports 35.57% accuracy in Table 1 but 37.57% in Table 2.

3 comments

W1. We sincerely thank the reviewer for the constructive comments. We address the three concerns regarding label availability, baseline comparisons, and communication efficiency as follows.

(1) Global sample indices and label availability.

We clarify that HOSFL does not require clients to upload labels. The server already has access to the label information and maintains the corresponding global sample-index-to-label mapping. Clients only upload intermediate representations and their global sample indices, allowing the server to associate each representation with its corresponding label. This setting is consistent with the original SplitFed framework [1], which explicitly considers label sharing and enables the Main Server to access ground-truth labels for loss computation. Thus, HOSFL does not introduce an additional centralized label source. We will further clarify this assumption in the revised manuscript.

(2) One-shot baseline comparisons and their applicability to SFL.

Existing generative and distillation-based one-shot FL methods cannot be directly transferred to SFL due to their reliance on complete local teacher models. As discussed in our Related Work, these methods typically require clients to sufficiently train complete local models before uploading them to the server for model ensembling, generation, or knowledge distillation. In contrast, SFL clients only possess partial front models and cannot independently obtain complete classification models. Under a strict one-shot constraint, repeated end-to-end split optimization is unavailable, making it difficult to obtain reliable teacher models. Simply splitting existing one-shot FL models into front and tail components therefore cannot resolve this incompatibility. Making them applicable would require fundamentally redesigning their optimization and knowledge construction mechanisms, rather than simply transferring the original methods.

This is precisely the motivation for HOSFL. As discussed in our Introduction, one-shot SFL is practically meaningful in scenarios where repeated client-server interactions are constrained by communication resources and privacy requirements. Meanwhile, existing one-shot FL methods cannot directly address this setting due to their reliance on complete local models. HOSFL fills this gap by constructing global class-aware supervision directly from one-shot uploaded intermediate representations.

Furthermore, the one-shot results of conventional SFL methods in Table 1 are intended to demonstrate the performance degradation when multi-round methods are directly restricted to a single communication round, which motivates the development of HOSFL. We also clarify that one-shot refers to the communication protocol rather than a single optimization step. The local self-supervised learning, server-side generation, and subsequent client optimization are intrinsic components of HOSFL. We do not claim identical computational budgets to multi-round methods; our objective is to achieve effective learning with substantially reduced communication.

(3) Quantitative comparison of total communication costs.

We agree that fewer communication rounds do not necessarily imply lower total communication volume. Therefore, we additionally quantify the actual transmitted data, including client uploads, downloads, and server-to-server transfers.

MethodClient Upload (MB)Client Download (MB)Server Transfer (MB)Total (MB)
HOSFL13,119.9573.5542.0313,235.52
MultiSFL63,675.4156,872.870.00120,548.28
SFL98,489.8098,483.800.00196,973.60
FedFSA69,731.0459,078.380.00128,809.42
FedAvg67,101.6167,101.610.00134,203.22
SL98,399.5998,393.590.00196,793.18

HOSFL reduces the total communication volume by 89.02%, 93.28%, 89.72%, 90.14%, and 93.27% compared with MultiSFL, SFL, FedFSA, FedAvg, and SL, respectively. These results demonstrate that HOSFL reduces not only communication rounds but also the actual transmitted data volume.

Reference

[1] Thapa C, Arachchige P C M, Camtepe S, et al. Splitfed: When federated learning meets split learning[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2022, 36(8): 8485–8493.

W2.

We thank the reviewer for the constructive comments. We conducted additional experiments and further clarified the design rationale of HOSFL from three aspects.

(1) Cross-client feature alignment and global prototype construction.

To examine the compatibility of independently optimized client encoders, we conducted feature alignment diagnostics on CIFAR-10 with 10 clients and Dirichlet α=0.3\alpha=0.3, using 1,000 common probe samples.

Diagnostic metricResult
Mean relative parameter dispersion0.01076
Mean pairwise linear CKA0.95372
Minimum pairwise linear CKA0.91433
Mean same-sample feature cosine0.74456
Mean class-prototype cosine0.96787
Mean prototype relative L20.24935

All clients share the same encoder architecture and initialization and employ the same self-supervised objective. Furthermore, global class prototypes are constructed by aggregating uploaded representations belonging to the same class, rather than simply averaging local prototypes. The high CKA and prototype cosine similarities demonstrate substantial consistency in representation structures and class-level feature directions across clients, supporting global prototype construction and complementing the conditional analysis in Appendix B.2.

(2) Rationale of prototype perturbation.

Regarding Eq. (11), an equivalent expected objective does not necessarily imply identical optimization dynamics. Stochastic perturbations introduce variations in prototype targets at individual optimization steps, leading to different gradient realizations and potentially different optimization trajectories during finite-step, non-convex generator training. This exposes the generator to a neighborhood of class-conditioned feature targets rather than a single deterministic prototype. Thus, perturbation serves as a stochastic optimization mechanism rather than an explicit diversity-promoting regularizer in the expected objective.

(3) Synthetic generation quality and component-wise validation.

First, we evaluated synthetic feature distributions on CIFAR-100 using the same frozen reference encoder (ErefE_{\mathrm{ref}}) for real and synthetic samples. With 500 samples of each type per class, the mean class-wise Feature FID across 100 classes is 24.85±8.1124.85 \pm 8.11. We emphasize that HOSFL aims to generate task-oriented synthetic supervision rather than photorealistic images or exact reconstructions of original data.

Second, we conducted additional component-wise ablations to evaluate the contributions of the two generator constraints.

ConfigurationCIFAR-10 Accuracy (%)
Full HOSFL61.90
Without prototype feature alignment48.65
Without classifier guidance53.78

Removing prototype feature alignment reduces accuracy by 13.25 percentage points, while removing classifier guidance results in an 8.12-percentage-point decrease. These results demonstrate the complementary contributions of the two constraints: prototype alignment anchors generated representations to global class prototypes, while classifier guidance reinforces class-discriminative consistency.

Furthermore, Table 2 evaluates the overall contribution of the generation and subsequent refinement stages, while Fig. 5 examines the sensitivity to the corresponding Stage II loss weights. The additional ablations provide more fine-grained evidence of the individual contributions of generator supervision mechanisms.

Overall, these additional experiments and analyses support the feasibility of cross-client feature aggregation, clarify the role of prototype perturbation, and validate the contributions of the proposed synthetic supervision mechanisms.

Q1.

Clients upload intermediate representations together with their global sample indices, but do not upload labels. The servers use these indices to retrieve the corresponding labels from the server-side index-to-label mapping. This is consistent with the original SFL framework [1], where the server computes supervised loss and propagates gradients through the split layer, which requires access to the corresponding labels. Therefore, HOSFL does not introduce an additional client-side label transmission mechanism.

Q2.

Table 1 serves two complementary purposes: comparing methods under their original multi-round protocols and examining whether conventional iterative methods can maintain their performance when restricted to one communication round. For split-based comparisons, we use the same ResNet-18 architecture, split location, client data partitions, and evaluation protocol. FedFSA is also adapted to the smashed-feature space while preserving its feature alignment and server-side calibration mechanisms. The near-random accuracy of one-shot SFL baselines reflects their dependence on repeated client-server optimization, rather than a weakness of their original multi-round methods. This comparison demonstrates that simply restricting iterative SFL methods to one round is insufficient. Moreover, existing generative and distillation-based one-shot FL methods rely on sufficiently trained complete local models as teacher models. Such models are unavailable in SFL, where clients only possess partial front models. Directly splitting these methods cannot provide reliable teacher knowledge under a single interaction and would require redesigning their supervision mechanisms. This fundamental incompatibility motivates HOSFL.

Q3.

All clients share the same encoder architecture and initialization and use the same self-supervised objective. To empirically evaluate feature compatibility, we conducted additional diagnostics on CIFAR-10 with 10 clients and α=0.3\alpha=0.3, using 1,000 common probe samples.

Diagnostic metricResult
Mean relative parameter dispersion0.01076
Mean pairwise linear CKA0.95372
Minimum pairwise linear CKA0.91433
Mean class-prototype cosine similarity0.96787
Minimum class-prototype cosine similarity0.93732

These results demonstrate substantial consistency in client representation structures and class-prototype directions, providing empirical support for global prototype aggregation and complementing the analysis in Appendix B.2.

Q4.

We conducted additional component-wise ablation experiments on CIFAR-10 with 10 clients and α=0.3\alpha=0.3.

ConfigurationAccuracy (%)
Full HOSFL61.90
Without prototype feature alignment48.65
Without classifier guidance53.78

Removing prototype feature alignment reduces accuracy by 13.25 percentage points, while removing classifier guidance reduces it by 8.12 percentage points. These results demonstrate the complementary contributions of the two components: prototype alignment constrains generated representations toward global class prototypes, while classifier guidance reinforces class-discriminative consistency. The existing ablation study in Table 2 further evaluates the overall contribution of synthetic generation and second-stage refinement.

Q5.

We apologize for this typographical error. The CIFAR-100 accuracy at α=0.3\alpha=0.3 in Table 2 was incorrectly reported as 37.57%. The correct value is 35.57%, consistent with Table 1. We will correct Table 2 accordingly.

Reviewer_PPwgOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes HOSFL, a one-shot split federated learning framework that combines two-stage client-side self-supervised learning, server-side training on smashed features, prototype-guided conditional generation, and synthetic-data-assisted client refinement to alleviate class imbalance and missing-class issues. The HOSFL is tested in three widely-used classification dataset.

Critical strengths
  1. the paper is well-organized and easy to follow. 2.The empirical results in Table 1 show clear gains over directly adapted one-shot baselines, and the ablation study suggests that both the first-stage SSL and the server-side refinement contribute to the final performance.
Critical weaknesses
  1. the proposed HOSFL only tested in three simple classification tasks.
  2. The core contribution is not clear. What is the advance of HOSFL compared with existing work. If it is just a combination of existing techniques, the novelty might be limited.
  3. Tha analysis might be not hold. The analysis in Appendix B is mostly a dependency/sufficiency description, and the remaining propositions rely on strong assumptions such as bounded parameter dispersion, a Lipschitz classifier, and a prototype margin, which do not substantiate the central claims.
  4. Why HOSFL’s single-round accuracy being close to or even higher than multi-round baselines?
Questions for authors
  1. Why HOSFL’s single-round accuracy being close to or even higher than multi-round baselines?
  2. What is the core contribution of HOSFL, compared with existing techniques/methods
2 comments

Q1. The proposed HOSFL only tested in three simple classification tasks. We would like to clarify that image classification is one of the most widely adopted evaluation tasks in existing Federated Learning (FL) and Split Federated Learning (SFL) studies, where classification accuracy is commonly used to evaluate model performance under different data heterogeneity settings. Therefore, we follow this established evaluation protocol to facilitate a consistent and meaningful comparison with existing methods.

Moreover, the three datasets used in our experiments, CIFAR-10, SVHN, and CIFAR-100, exhibit different data characteristics, class distributions, and classification complexities, providing complementary evaluations of HOSFL under heterogeneous data conditions. In particular, CIFAR-100 introduces a more challenging classification setting with a larger number of classes. We agree that extending HOSFL to other tasks, such as object detection and semantic segmentation, would be valuable for further investigating its applicability. We consider this an important direction for future work.

Q2. The theoretical analysis in Appendix B relies on strong assumptions and may not sufficiently support the central claims of HOSFL.

We would like to clarify the intended scope and role of the theoretical analysis in Appendix B. The analysis is designed to provide theoretical insights into the key components of HOSFL, rather than establish an end-to-end convergence or generalization guarantee. Specifically, Proposition B.1 establishes the sufficiency of the information collected after one-shot communication for subsequent server-side generator training, explaining the feasibility of the proposed communication protocol. Proposition 1 characterizes how client-side parameter dispersion affects the discrepancy between client-derived prototypes and the reference encoder space, where the dispersion explicitly determines the tightness of the derived bound rather than requiring identical client models. Similarly, the result in B.4 provides a sufficient condition connecting feature proximity to class prototypes with semantic consistency under the stated Lipschitz and prototype-margin assumptions. These results characterize specific properties of our design under clearly defined conditions, rather than making unconditional guarantees.

Importantly, the central claims of HOSFL, including its effectiveness under heterogeneous and class-imbalanced data distributions, are further supported by extensive experiments across three datasets, different heterogeneity levels, and comparisons with existing methods. Thus, Appendix B provides theoretical justification for the individual design components, while the experimental results demonstrate the effectiveness of the complete framework.

Q3. What is the core contribution of HOSFL, compared with existing techniques/methods

We would like to clarify that the motivation for HOSFL is not simply to combine existing techniques, but to address a practically important yet underexplored problem in split federated learning. In real-world resource-constrained environments, such as edge computing and intelligent devices, repeated client-server interactions incur considerable communication and computational costs, making conventional multi-round SFL less suitable for scenarios with strict communication budgets. Therefore, investigating one-shot SFL is practically meaningful: it aims to enable effective collaborative learning when only a single round of client-server interaction is affordable, rather than merely reducing communication rounds for the sake of efficiency.

The key challenge is that existing one-shot FL methods generally rely on complete local models to provide global knowledge, whereas clients in SFL only possess partial front models and cannot independently provide complete classification supervision. To address this limitation, HOSFL develops a split-compatible global supervision construction mechanism using only one-shot uploaded information. Specifically, it combines cross-client class prototypes, prototype-guided feature alignment, and classifier-guided class consistency to generate class-balanced synthetic supervision. This enables the server to recover both class-level representation information and discriminative supervision without additional client-server interactions. Furthermore, a two-stage client refinement strategy incorporates this supervision while preserving real-data-driven representation learning and stability. The main advance of HOSFL is a unified framework that makes effective one-shot collaborative optimization feasible under the structural constraints of SFL, particularly in the presence of severe class imbalance and missing classes. The experimental results and ablation studies further demonstrate its effectiveness compared with existing SFL baselines.

Q4. Why HOSFL’s single-round accuracy being close to or even higher than multi-round baselines

The competitive performance of HOSFL stems from its targeted treatment of class imbalance rather than the number of communication rounds alone. FedAvg, SL, and SFL are general collaborative learning frameworks, while MultiSFL primarily addresses general data heterogeneity through multi-model aggregation and knowledge replay. None of these methods explicitly targets severe class imbalance and missing-class scenarios.

This is an important yet insufficiently explored problem in SFL. Existing studies have mainly focused on system-level heterogeneity, such as communication and computational efficiency, while client-side class imbalance has received much less attention. In practical non-IID settings, clients may have highly fragmented label distributions or even completely missing classes, which cannot necessarily be resolved by simply increasing training rounds.

FedFSA is specifically designed for class imbalance, but its original formulation assumes complete local models. We therefore adapt it to the SplitFed architecture by performing feature alignment in the smashed-feature space while retaining server-side calibration, ensuring a fair comparison under the same model partition.

Further evidence is provided in Appendix C.2, where clients have mutually disjoint class subsets. HOSFL achieves 46.33%, 33.25%, and 27.89% accuracy on SVHN, CIFAR-10, and CIFAR-100, respectively, consistently outperforming the other compared methods.

Reviewer_kz5POpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates one-shot split federated learning (SFL), where clients hold only the front part of the network, interact with the server only once, and face non-IID, class-imbalanced data including missing-class scenarios. The proposed HOSFL proceeds in stages: clients first train front models self-supervised with a BYOL-style objective enriched by multi-crop local-view alignment and a VICReg-style covariance term, then upload smashed features once. A main server trains a shared tail model on these features with temperature-scaled cross-entropy and logit regularization. A separate federated server aggregates cross-client class prototypes, perturbs them for intra-class diversity, and trains a conditional generator with a prototype feature-alignment loss plus a classifier-guided consistency loss. Class-balanced synthetic samples are then sent back to clients for a second-stage refinement with a frozen-encoder feature regularizer and a classification term on synthetic data. On CIFAR-10/100 and SVHN with ten clients, HOSFL with one round reportedly reaches 86.76-91.62% on SVHN and 46.87-67.67% on CIFAR-10, while multi-round baselines (FedAvg, SL, SFL, MultiSFL, FedFSA) restricted to one shot collapse to near-random accuracy; HOSFL even surpasses the 150-round versions of these baselines, with an average gain of 4.72%.

Critical strengths
  • A novel and practically relevant problem. To my knowledge this is the first work to study one-shot communication in the SFL setting, and the paper correctly identifies that existing one-shot FL methods assume complete local models — an assumption that fails under splitting. The joint treatment of one-shot communication and class imbalance is well motivated, and this problem-framing contribution is the main factor in my assessment.
  • An ingeniously assembled framework with striking results. The combination of self-supervised front training, cross-client prototype aggregation, perturbed-prototype feature alignment, and classifier-guided generation is thoughtfully put together, and the reported margins, one-shot HOSFL surpassing 150-round baselines would be very significant if fully substantiated.
Critical weaknesses
  • The privacy posture is unclear and possibly self-defeating. Section 3.1 states that uploaded sample indices are used on the server side to recover class labels, meaning labels effectively leave the client — undermining a large part of split learning's privacy motivation. The two-server architecture (main server + federated server) implicitly assumes non-collusion but never states the threat model, and the differential privacy claim ("applies the same differential privacy noise as SFL") comes with no privacy budget, noise scale, formal guarantee, or any with/without-noise experiment. For a framework whose uploads include features, front-model parameters, and label-recoverable indices, this is a material gap.
  • The one-shot and efficiency accounting is incomplete, and the comparison is somewhat strawman. The system involves multiple transmissions — features and front models up, tail model between servers, synthetic datasets back down to every client — yet the communication-overhead claim is supported only by a time-complexity figure, with no megabyte-level budget that includes distributing class-balanced synthetic data (ratio_fake=0.5). Meanwhile all one-shot baselines are multi-round methods with communication simply truncated; none was designed for one round, so the 70+-point gaps invite scrutiny, and no dedicated one-shot generative method (e.g., DENSE-style) is adapted for comparison.
Questions for authors
  • [Clarification] Exactly what label information does each server obtain? If sample indices let the server recover ground-truth labels, how does this square with the privacy motivation of split learning, and can the method work when clients hold labels privately?
  • [Clarification] What is the assumed threat model between the main server and the federated server? What breaks if they collude, given one holds features and front models while the other holds prototypes, generator, and tail model?
  • [Clarification] Please provide the DP analysis: epsilon, noise scale, where noise is applied, and an accuracy-privacy trade-off curve — or tone down the claim.
  • [Clarification] Please report the full communication budget in megabytes, including the synthetic datasets distributed to clients, against the baselines.
  • [Clarification] Can you adapt a dedicated one-shot FL method (generative or distillation-based) to the split setting for a fairer one-shot comparison?
  • [Clarification] What evidence supports generator fidelity on CIFAR-100 from smashed features alone (e.g., FID, or t-SNE of synthetic vs. real features)?
  • [Minor] Please unify notation (E_ref is used only for feature extraction yet is described as "aggregating client-side front model parameters"); Figure 3's caption mentions "three diverse alpha values" while the text lists alpha in {0.3, 0.1}; please hedge the "first work" claim appropriately.
3 comments

We thank the reviewer for the constructive comments. We provide the following point-by-point clarifications.

1、Server-side label access. In HOSFL, clients upload intermediate representations with their global sample indices, allowing both servers to retrieve the corresponding labels through the index-to-label mapping. The Main Server uses these labels to calculate the cross-entropy loss for training the shared Tail Model (Eq. 8), while the Federated Server uses them to group representations and construct global class prototypes (Eq. 10). Thus, both servers can access the labels associated with uploaded representations, but neither accesses clients' raw images.

SFL privacy assumptions. This design follows the label-sharing configuration of conventional SplitFed Learning (SFL) [1]. Specifically, Algorithm 1 in [1] describes clients transmitting smashed data together with corresponding labels to the Main Server for loss calculation and backpropagation. The primary privacy objective of split learning is to keep raw data on clients, rather than inherently guarantee label confidentiality. HOSFL adopts this configuration, and we will clarify the distinction between raw-data privacy and label privacy in the revised manuscript.

Strictly label-private settings. We acknowledge that HOSFL is not directly applicable when sample-level labels must remain private. Potential extensions include: (1) moving label-dependent loss computation to clients, which return the corresponding gradients to the server; and (2) using secure aggregation of class-wise feature sums and sample counts to construct global prototypes without exposing individual labels. MPC or HE could also support label-dependent operations without revealing plaintext labels. These extensions require further investigation and may introduce additional computational and communication overhead, particularly in the one-shot setting.

[1]Thapa C, Arachchige P C M, Camtepe S, et al. Splitfed: When federated learning meets split learning[C]//Proceedings of the AAAI conference on artificial intelligence. 2022, 36(8): 8485-8493.

2、Assumed threat model. HOSFL follows the conventional two-server SFL architecture [1], consisting of clients, a Main Server, and a Federated Server. The two servers undertake different functional responsibilities during training. We assume that both servers follow the prescribed protocol but may analyze the information available to them (honest-but-curious), and are independently operated without collusion. Specifically, the Main Server receives intermediate features and corresponding sample indices, and trains the shared Tail Model using the associated labels. The Federated Server receives intermediate features and client-side front-model parameters, constructs global class prototypes, and aggregates the front models to obtain a reference encoder. After receiving the updated Tail Model parameters from the Main Server, the Federated Server combines the reference encoder and Tail Model to construct a global classifier for generator training.

Potential impact of server collusion. If the two servers collude, they can jointly analyze the information available across both server roles, potentially increasing the risk of inferring clients' original training data. However, the combined server-side information does not necessarily reveal the original samples directly. HOSFL retains raw training data locally and applies the same differential privacy noise mechanism as conventional SFL during intermediate feature transmission. Moreover, its one-shot communication paradigm avoids repeated feature exchanges. It is worth noting that the Federated Server already utilizes the intermediate features, class labels, front-model parameters, and updated Tail Model parameters to train the generator. Specifically, the generator is optimized using two complementary objectives: prototype-guided feature alignment, which encourages generated representations to match the corresponding class prototypes, and classifier-guided class consistency, which ensures that generated samples maintain their target class identities. Thus, the generator learns class-level characteristics from the available feature and model information and produces synthetic samples for clients' subsequent training. As shown in Fig. 7, although the generated samples preserve recognizable class characteristics, they exhibit clear differences from the original training images, without apparent reproduction of their instance-specific structures and details. This provides empirical evidence that the generator does not simply memorize or reproduce the original client samples, even when trained using the combined feature and model information.

3、HOSFL adopts the PixelDP-based feature perturbation mechanism introduced in conventional SFL [1]. Specifically, a noise layer is incorporated into the client-side network to perturb intermediate activations before they are transmitted to the server. Following [1], the perturbed activation is expressed as Ak,iP=Ak,i+Lap⁡(Δp,qAϵ′)A_{k,i}^{P}=A_{k,i}+\operatorname{Lap}\left(\frac{\Delta_{p,q}^{A}}{\epsilon'}\right), where Δp,qA\Delta_{p,q}^{A} denotes the sensitivity of the intermediate activation, and ϵ′\epsilon' is the privacy budget controlling the Laplace noise scale. Thus, the noise scale is Δp,qA/ϵ′\Delta_{p,q}^{A}/\epsilon'. A smaller ϵ′\epsilon' introduces stronger perturbation, providing stronger privacy protection at the potential cost of model utility.

The original SplitFed study further illustrates this privacy–utility relationship by evaluating different values of ϵ′\epsilon' (0.5, 1, 2, and 5), showing that model accuracy generally increases as the perturbation becomes weaker [1].

We acknowledge that the current HOSFL experiments do not include a dedicated accuracy–privacy trade-off curve across different privacy budgets. Our existing PSNR and SSIM evaluations instead examine the reconstruction quality of transmitted intermediate features under the adopted perturbation setting. We will clarify the mechanism, its application location, and the scope of the existing privacy evaluation in the revised manuscript, and avoid overstating the privacy guarantee beyond the evaluated setting.

4、We agree that reporting the complete communication budget is necessary for a fair comparison, particularly because HOSFL introduces an additional synthetic dataset distribution stage.

To address this concern, we calculated the communication overhead of HOSFL and all baseline methods over their complete training processes. The communication cost includes client uploads, client downloads, and inter-server transmissions, where applicable. Importantly, the communication cost of HOSFL explicitly accounts for both the initial model distribution and the synthetic datasets delivered to clients. All results are reported in decimal megabytes (MB).

The complete communication budget is summarized in Table X.

MethodClient Upload (MB)Client Download (MB)Server Transfer (MB)Total (MB)
HOSFL13,119.9573.5542.0313,235.52
MultiSFL63,675.4156,872.870.00120,548.28
SFL98,489.8098,483.800.00196,973.60
FedFSA69,731.0459,078.380.00128,809.42
FL (FedAvg)67,101.6167,101.610.00134,203.22
SL98,399.5998,393.590.00196,793.18

As shown in Table X, HOSFL incurs a total communication cost of 13,235.52 MB, including the transmission of synthetic datasets to clients. Although HOSFL introduces this additional data distribution stage, its overall communication overhead remains lower than those of the compared baselines under their respective complete training configurations.

These results demonstrate that the additional communication introduced by synthetic data distribution is limited compared with the repeated model exchanges or intermediate-feature transmissions required by the baseline methods. We will include this complete communication analysis in the revised manuscript.

5、We agree that generative and distillation-based one-shot FL methods are relevant to our work. However, directly adapting them to SFL is non-trivial due to a fundamental difference in their assumptions. Existing one-shot FL methods typically rely on sufficiently optimized complete local models, which serve as reliable teachers for server-side model ensembling, generation, or knowledge distillation. In contrast, SFL clients only possess partial front models and cannot independently produce complete classification predictions. Simply splitting these existing models into front and tail components does not resolve this issue. Under a strict one-shot constraint, repeated end-to-end split optimization is unavailable, making it difficult to obtain reliable teacher models from partially optimized front models and a shared server-side tail. Achieving effective adaptation would therefore require fundamentally redesigning their knowledge construction and supervision mechanisms, rather than simply transferring the original methods to SFL. This incompatibility motivates HOSFL. Instead of relying on complete local teacher models, HOSFL constructs global class-aware supervision directly from one-shot uploaded intermediate representations through prototype aggregation, prototype-guided feature alignment, and classifier-guided class consistency. Meanwhile, the one-shot experiments in Table 1 serve a complementary purpose: they demonstrate that conventional multi-round SFL methods cannot maintain their performance simply by restricting communication to a single round.

6、To further evaluate the generator on CIFAR-100, we conducted additional quantitative and qualitative feature-level evaluations. Specifically, real and synthetic features are extracted using the same frozen reference encoder (ErefE_{\mathrm{ref}}), ensuring a consistent feature space. Using 500 real and 500 synthetic samples per class, the mean class-wise Feature FID across all 100 classes is 24.85±8.1124.85 \pm 8.11.

We further provide a t-SNE visualization of synthetic features from 20 representative classes. The visualization shows distinct class-dependent clusters, indicating that the generator captures meaningful class-related structures and maintains class separability. Meanwhile, the Feature FID provides a quantitative assessment of the distribution differences between real and synthetic features.

We emphasize that HOSFL aims to generate class-consistent synthetic supervision rather than exactly reconstruct real samples or reproduce the entire real feature distributions. These additional results provide complementary quantitative and qualitative evidence of the generator's ability to capture class-related feature structures on CIFAR-100 using the information available after one-shot communication. We have added the corresponding results and discussion to the revised manuscript.

7、We have carefully revised the corresponding parts of the manuscript, as detailed below.

**Clarification of ErefE_{\mathrm{ref}}: ** We have clarified that ErefE_{\mathrm{ref}} is constructed by applying FedAvg to the front-model parameters uploaded by the ten clients. During generator training, it remains frozen and serves as a reference encoder for extracting features from generated images. The extracted features are then aligned with the corresponding global class prototypes through the feature-matching loss LfmL_{fm}. We have revised the description to distinguish the parameter aggregation process from the feature extraction function of ErefE_{\mathrm{ref}}.

Clarification of the heterogeneity levels in Fig. 3: Our experiments consider three Dirichlet heterogeneity levels, α∈{0.1,0.3,0.5}\alpha \in \{0.1, 0.3, 0.5\}. However, for better visual clarity and to avoid overlapping distributions, Fig. 3 only illustrates two representative levels, α=0.1\alpha=0.1 and α=0.3\alpha=0.3, with the less heterogeneous case (α=0.5\alpha=0.5) omitted from the visualization. The original figure caption was not updated accordingly. We have corrected the caption to accurately reflect the two displayed levels and clarified in the text that all three levels are included in the experiments.

Refinement of the novelty claim: To the best of our knowledge, we have not identified prior work explicitly investigating one-shot communication in the SFL setting. Nevertheless, to avoid an overly absolute priority claim, we have refined the original statement to emphasize that HOSFL provides an initial study of this challenging setting.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.