- 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
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 , using 1,000 common probe samples.
| Diagnostic metric | Result |
|---|---|
| Mean relative parameter dispersion | 0.01076 |
| Mean pairwise linear CKA | 0.95372 |
| Minimum pairwise linear CKA | 0.91433 |
| Mean same-sample feature cosine | 0.74456 |
| Mean class-prototype cosine | 0.96787 |
| Mean prototype relative L2 | 0.24935 |
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 () for real and synthetic samples. With 500 samples of each type per class, the mean class-wise Feature FID across 100 classes is . 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.
| Configuration | CIFAR-10 Accuracy (%) |
|---|---|
| Full HOSFL | 61.90 |
| Without prototype feature alignment | 48.65 |
| Without classifier guidance | 53.78 |
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 , using 1,000 common probe samples.
| Diagnostic metric | Result |
|---|---|
| Mean relative parameter dispersion | 0.01076 |
| Mean pairwise linear CKA | 0.95372 |
| Minimum pairwise linear CKA | 0.91433 |
| Mean class-prototype cosine similarity | 0.96787 |
| Minimum class-prototype cosine similarity | 0.93732 |
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 .
| Configuration | Accuracy (%) |
|---|---|
| Full HOSFL | 61.90 |
| Without prototype feature alignment | 48.65 |
| Without classifier guidance | 53.78 |
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 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.
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.
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.