Pith. sign in

REVIEW 4 major objections 4 minor 2 cited by

GAPSL argues that the convergence collapse of aggregation-free parallel split learning under non-IID client data is caused by inter-device gradient directional inconsistency, and that a dynamically selected leader gradient plus a cosine ali

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 05:45 UTC pith:3LWG3S3E

load-bearing objection Real problem, solid LGI story, but the GDA regularizer as implemented is mathematically inert — the core alignment claim needs a fix before the paper can be taken at face value. the 4 major comments →

arxiv 2603.18540 v2 pith:3LWG3S3E submitted 2026-03-19 cs.LG

GAPSL: A Gradient-Aligned Parallel Split Learning over Data-Heterogeneous Edge Computing Systems

Zheng Lin , Ons Aouedi , Zihan Fang , Wei Ni , Yue Gao , Symeon Chatzinotas , Xianhao Chen This is my paper
classification cs.LG
keywords parallel split learninggradient directional inconsistencydata heterogeneityleader gradient identificationgradient direction alignmentnon-IID dataedge computingdistributed machine learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes GAPSL, a parallel split learning framework that keeps the communication and deployment advantages of aggregation-free PSL while fixing its tendency to diverge when client data is unevenly distributed. The central idea is to have the server identify a leader gradient from the client gradients that point most consistently in the same direction, then pull each client's update toward that leader with a cosine-based penalty. If the mechanism works as claimed, aggregation-free split learning can remain stable under severe data heterogeneity and even beat split federated learning in accuracy and convergence speed. The paper reports non-IID gains of up to 16.0 accuracy points over PSL on CIFAR-10 and 13.8 points on CIFAR-100.

Core claim

GAPSL's central claim is that the divergence of aggregation-free PSL under heterogeneous data is a gradient-direction problem, not a fundamental limitation of split learning. The server computes each client's gradient on its own segment, scores each client by the average angular deviation from every other client's gradient, selects a top fraction whose size adapts to dispersion and training round, and averages the survivors into a 'leader gradient' that approximates the global descent direction. Then, for clients within an adaptive angular threshold of the leader, the loss is augmented with λ(1−cos θ_lead,i), a cosine penalty that vanishes when aligned. Evaluated on CIFAR-10/VGG-16 and CIFAR

What carries the argument

Two coupled server-side mechanisms. Leader Gradient Identification (LGI) computes pairwise angular deviations arccos(⟨g_i,g_j⟩/(‖g_i‖‖g_j‖)), averages them into a per-client consistency score, selects the top K_t% of clients with K_t = K_min + (t/T)·(relative stability score)·(K_max−K_min), and averages their server-side gradients into g_lead. Gradient Direction Alignment (GDA) then computes each client's angular deviation θ_lead,i from g_lead, builds an adaptive threshold max(min(μ−ησ, π/2), 0) from the mean and standard deviation of those deviations, filters out clients beyond the threshold, and adds λ(1−cos θ_lead,i) to the surviving clients' losses. The cosine penalty is intended to vani

Load-bearing premise

The load-bearing premise is the paper's footnote to Eq. (7), that treating the client and leader gradient vectors as detached constants preserves the GDA regularizer's alignment effect; if a detached constant is used, the penalty has zero gradient, so the described alignment mechanism would not exist in the evaluated system and any gains would have to come from LGI selection and threshold filtering alone.

What would settle it

Inspect the evaluated implementation's backward pass and compute the gradient of λ(1−cos θ_lead,i) with respect to the server-side model parameters while g_i and g_lead are detached; if that gradient is numerically zero, the GDA term changes no update, and any accuracy difference between LGI+GDA and LGI+Non-GDA must come from the filtering or averaging strategy rather than from the claimed alignment penalty.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If GAPSL is correct, aggregation-free PSL can remain stable under severe non-IID data, preserving PSL's lower communication and deployment costs relative to SFL.
  • The same server-side coordination applies without any client-side model sharing, so the privacy properties of split learning are retained.
  • Reported non-IID accuracy gains over PSL are up to 16.0 points on CIFAR-10 and 13.8 points on CIFAR-100, with faster convergence on both datasets and models.
  • The adaptive selection ratio and angular threshold remove the need to tune a fixed selection fraction or threshold for each task.
  • The method maintains accuracy better than PSL as the number of devices grows to 40 under non-IID data, suggesting scalability to larger edge systems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the footnote to Eq. (7) is taken literally, that g_i and g_lead are detached constants, then the GDA penalty contributes zero gradient to the update, so the measured GAPSL-vs-PSL gains would be explained by LGI's selective averaging and threshold filtering rather than by the direction-aware alignment mechanism described in Section III-C.
  • A direct test would be to implement the regularizer without detachment, using true double backpropagation; if accuracy rises, the original first-order implementation was leaving the paper's own mechanism unused, and if it does not, the stability comes from client selection, not alignment.
  • The adaptive threshold mean−η·std assumes the angular-deviation distribution is roughly symmetric and bell-shaped; a rank-based or percentile threshold would be a natural alternative if the distribution is skewed, as is often the case under extreme class imbalance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes GAPSL, a parallel split learning (PSL) framework designed to mitigate the training divergence caused by non-IID client data. GAPSL has two components: leader gradient identification (LGI), which computes pairwise angular deviations between client gradients, adaptively selects the top-K most directionally consistent gradients, and averages them into a leader gradient; and gradient direction alignment (GDA), which computes each client's angular deviation to the leader gradient, filters out clients whose deviation exceeds an adaptive threshold based on mean−η·std, and adds a λ(1−cosθ) regularization term to the losses of the remaining clients. The method is evaluated on a Jetson-based testbed with CIFAR-10/VGG-16 and CIFAR-100/ViT-Base under IID and non-IID settings. The paper reports consistent improvements in accuracy and convergence time over PSL, SFL, EPSL, and vanilla SL, including non-IID gains of up to 16.0 points over PSL on CIFAR-10.

Significance. If the method worked as described, it would be a valuable contribution to aggregation-free split learning: it removes client-side model aggregation while addressing gradient conflict, and the hardware testbed evaluation with two model families is more realistic than simulation-only studies. The empirical motivation in Section II is clearly presented, and the ablation studies are a useful attempt to isolate LGI and GDA. However, the central mechanism is undermined by the paper's own Footnote 6, and several hyperparameters are selected on the same datasets used for evaluation, without independent validation or variance reporting. As written, the reported results do not establish the claimed direction-alignment effect.

major comments (4)
  1. [§III-C3, Eq. (7) and Footnote 6] The central mechanism is not present in the evaluated system. Footnote 6 states that the implementation treats g_i and g_lead as detached constants from the computational graph. Under that implementation, θ_lead,i depends only on detached tensors, so ∂λ(1−cos θ_lead,i)/∂w = 0 for every trainable parameter w. The regularization term contributes exactly zero gradient, and Eq. (7) reduces to L_i plus a constant. Therefore the ablation in Fig. 21 (LGI+GDA vs. LGI+Non-GDA) cannot be explained by the proposed GDA; the only active component is the adaptive threshold of Eq. (6), which is client selection, not alignment. The authors must either document the actual differentiable path for Eq. (7) or withdraw the alignment claim.
  2. [§III-C2, Eq. (6)] The threshold θ_th = max(min(μ − ην, π/2), 0) is presented as a 'robust confidence boundary' without any distributional justification. For skewed or multi-modal angular deviations, mean−η·std is not necessarily a meaningful outlier cutoff, and the clamps at 0 and π/2 can discard all clients or admit all clients in ways that are not analyzed. Moreover, η is not reported in Section IV-A4 and is not ablated in Section V-C. Since this threshold is the only remaining part of GDA that can affect training once Eq. (7) is inert, its unspecified handling is a concrete gap.
  3. [§IV-A4 and §V] The hyperparameter protocol is under-specified. λ is set to 5×10−4 for VGG-16 and 3×10−5 for ViT-Base, K_min/K_max to 20%/80%, and η is never stated. These values appear to be chosen on CIFAR-10/CIFAR-100, and Tables I–III report single runs without standard deviations or significance tests. Without a described model-selection procedure (hold-out validation, grid ranges, repeated seeds), the claimed consistent gains may reflect tuning to the test sets rather than a robust algorithmic advantage.
  4. [§III-B2, Eq. (2)] The adaptive selection ratio is justified only by the empirical observations in Fig. 10, which come from the same CIFAR-10/VGG-16 setup. The linear t/T term and the multiplicative coupling with the normalized dispersion term are ad hoc; no theoretical derivation or comparison with alternative functional forms is given. This weakens the claim that LGI's adaptivity is principled rather than curve-fit to the observed datasets.
minor comments (4)
  1. [§III-A] Typo: 'deivce' should be 'device'. Also, the sentence 'PSL incurs the same communication overhead as standard PSL' is tautological and should be rephrased.
  2. [§IV-A4] Duplicate wording: 'The gradient selection ratio is is bounded...' Also, the regularization coefficient λ is defined twice in the same paragraph.
  3. [Fig. 20(b)] The legend appears to omit the LGI+GDA curve, making it impossible to identify which line corresponds to the full method. Please fix the figure.
  4. [§III-B1, Eq. (1)] The symbol S is used both as the set of client devices and as the number of devices (|S|), which is slightly confusing. Consider using a different symbol for the set, such as C.

Circularity Check

1 steps flagged

Footnote 6 makes the GDA regularizer's gradient identically zero, so Eq. (7) is a no-op and the claimed alignment mechanism is vacuous by construction.

specific steps
  1. other [Section III-C3, Eq. (7), and Footnote 6 (with implementation in Section IV-A)]
    "6Theoretically, optimizing Eqn. (7) differentiating a loss function that incorporates the local gradient g𝑡𝑖, which necessitates higher-order differentiation (i.e., double backpropagation). To preserve the lightweight nature of GAPSL, our implementation adopts a pragmatic first-order approximation by treating the gradient vectors g𝑡𝑖 and g𝑡lead as detached constants from the computational graph."

    The advertised alignment force is the gradient of the added term λ(1−cosθ_lead,i) in Eq. (7). Because g_i and g_lead are detached, θ_lead,i is constant w.r.t. every trainable parameter, so ∂λ(1−cosθ_lead,i)/∂w = 0 identically. The regularized loss therefore has the same gradient as the plain local loss; GDA's regularization cannot alter any update. Fig. 21's claimed GDA gains are thus not produced by Eq. (7). Whatever remains is adaptive thresholding (client selection), not direction-aware alignment, so the central GDA claim reduces to a no-op by construction.

full rationale

No conventional circularity via self-citation or fitted inputs is present: LGI's leader gradient being built from the same server-side gradients it guides is a design property, not a derivation loop; the hyperparameters (λ, K_min/K_max, η) are tuned heuristic choices, not predictions generated from the model; and the many self-citations are background context, not load-bearing uniqueness theorems. However, the paper's own Footnote 6 creates a by-construction nullification of the GDA regularizer: the term claimed to perform direction alignment has zero derivative, so the central mechanism described in Section III-C cannot be what drives the reported results. This is a partial-vacuity/circularity of the claimed mechanism rather than a mere tuning concern, hence the non-zero score.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 1 invented entities

The method rests on several hand-tuned components (λ, K bounds, η) and on the empirical validity of the leader-gradient construction. There is no new physical or formal entity; the only invented algorithmic entity is the self-referential leader gradient. The most consequential premise — that detached gradients preserve the regularizer's effect — is false, and a second (Gaussian-style thresholding) is unvalidated.

free parameters (3)
  • λ (regularization coefficient) = 5×10⁻⁴ (VGG-16); 3×10⁻⁵ (ViT-Base)
    Chosen per model/dataset in Section IV-A4; directly controls the strength of the GDA regularizer that the paper claims is responsible for gradient alignment.
  • K_min / K_max (selection ratio bounds) = 20% / 80%
    Hand-set in Section IV-A4; bound the adaptive selection ratio in Eq. (2) and therefore the composition of the leader gradient.
  • η (angular threshold sensitivity) = not specified
    Controls the adaptive angular threshold in Eq. (6); no value is given anywhere in the paper, even though it directly determines which client gradients are excluded from training.
axioms (5)
  • domain assumption Angular deviation between server-side gradients is a faithful proxy for client-side data divergence and for the global convergence direction.
    This motivates LGI in Section III-B1 and is validated only indirectly through accuracy; no theoretical or statistical argument is provided.
  • ad hoc to paper The mean−η·std threshold of Eq. (6) validly separates informative from destructive gradients.
    No distributional model is stated for the angular deviation values; the paper critiques fixed thresholds but does not justify why this particular mean/std confidence boundary is robust.
  • domain assumption Uniformly averaging the top-K selected gradients yields a robust leader gradient; gradient magnitudes can be ignored.
    Eq. (4) discards magnitudes; the paper argues that small consistent gradients are more informative, but offers no analysis of when this holds or how outliers in magnitude affect the average.
  • ad hoc to paper The optimal selection ratio grows linearly with the round index (t/T in Eq. 2).
    Motivated by a single empirical sweep on one dataset (Fig. 10b); no derivation is given, and K_t directly changes which gradients form the leader.
  • ad hoc to paper Detaching the gradient vectors in the GDA regularizer does not nullify the alignment effect.
    Footnote 6; this premise is false as stated, since the derivative of a constant term with respect to model parameters is identically zero.
invented entities (1)
  • leader gradient (g_lead) no independent evidence
    purpose: Directional anchor to which all client gradients are aligned in GDA; also drives the client-selection threshold.
    Constructed from the same set of gradients it is used to align; it has no externally falsifiable handle outside the paper. If the no-op regularizer is real, the leader gradient only affects client selection, not the update direction.

pith-pipeline@v1.3.0-alltime-deepseek · 174 in / 19339 out tokens · 217273 ms · 2026-08-04T05:45:39.454590+00:00 · methodology

0 comments
read the original abstract

The increasing complexity of neural networks poses significant challenges for democratizing federated learning (FL) on resource-constrained edge devices. Parallel split learning (PSL) has emerged as a promising solution by offloading substantial computing workload to a server via model partitioning, shrinking client-side computing load, and eliminating the client-side model aggregation for reduced communication and deployment costs. However, the highly heterogeneous nature of client data in edge computing systems causes aggregation-free PSL to suffer from severe training divergence, stemming from gradient directional inconsistency across clients. To address this challenge, we propose GAPSL, a gradient-aligned PSL framework tailored for data-heterogeneous edge systems, which comprises two key components: leader gradient identification (LGI) and gradient direction alignment (GDA). LGI dynamically selects a set of directionally consistent device gradients to construct a leader gradient as a robust proxy for the global convergence trend. GDA employs a direction-aware regularization to align each client's gradient with the leader gradient, thereby mitigating inter-device gradient directional inconsistency and enhancing model convergence. We evaluate GAPSL on a prototype computing testbed. Extensive experiments demonstrate that GAPSL consistently outperforms state-of-the-art benchmarks in training accuracy, convergence latency, and system robustness under severe data heterogeneity.

Figures

Figures reproduced from arXiv: 2603.18540 by Ons Aouedi, Symeon Chatzinotas, Wei Ni, Xianhao Chen, Yue Gao, Zheng Lin, Zihan Fang.

Figure 1
Figure 1. Figure 1: The comparison of SFL, PSL, and GAPSL frameworks, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The training performance (a) and test accuracy (b) of [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: The PSL performance of training a model with the [PITH_FULL_IMAGE:figures/full_fig_p003_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: An overview of GAPSL architecture. B. Gradient Directional Inconsistency While Section II-A underscores that the lack of client-side model aggregation deteriorates the training performance of PSL, the underlying cause stems from inconsistent gradient directions across client devices. In PSL, each device independently performs forward propagation (FP) and BP on its local dataset, without any inter-device ag… view at source ↗
Figure 7
Figure 7. Figure 7: The normalized gradient consistency scores and corre [PITH_FULL_IMAGE:figures/full_fig_p004_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The test accuracy versus gradient selection ratio and [PITH_FULL_IMAGE:figures/full_fig_p004_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: The test accuracy versus STD of gradient consistency [PITH_FULL_IMAGE:figures/full_fig_p005_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: The test accuracy and converged time via leader [PITH_FULL_IMAGE:figures/full_fig_p006_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: An illustration of GDA. the global optimization trend. Empirical validation in [PITH_FULL_IMAGE:figures/full_fig_p006_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: The test accuracy versus angular threshold and epochs [PITH_FULL_IMAGE:figures/full_fig_p007_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: GAPSL prototype and testbed. 1) Dataset: To evaluate the learning performance of GAPSL, we adopt two widely used image classification datasets: CIFAR￾10 and CIFAR-100 [33]. CIFAR-10 consists of 10 distinct object classes with 50,000 training and 10,000 test images, while CIFAR-100 comprises 100 fine-grained categories, with each category containing 500 training and 100 test samples. In our experiment, we … view at source ↗
Figure 16
Figure 16. Figure 16: The converged accuracy for CIFAR-10 and CIFAR-100 [PITH_FULL_IMAGE:figures/full_fig_p009_16.png] view at source ↗
Figure 15
Figure 15. Figure 15: The training performance for CIFAR-10 and CIFAR [PITH_FULL_IMAGE:figures/full_fig_p009_15.png] view at source ↗
Figure 19
Figure 19. Figure 19: The test accuracy versus degrees of data heterogeneity [PITH_FULL_IMAGE:figures/full_fig_p010_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: The ablation evaluation for LGI on CIFAR-10 and [PITH_FULL_IMAGE:figures/full_fig_p011_20.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. FluxShard: Motion-Aware Feature Cache Reuse for Collaborative Video Analytics in Mobile Edge Computing

    cs.NI 2026-05 unverdicted novelty 7.0

    FluxShard uses per-block motion vectors and a Receptive Field Alignment Principle to manage feature cache reuse in edge-cloud video analytics, delivering 32.6-83.8% lower latency and 14.9-64.0% lower energy than basel...

  2. SwarmSense-DNN: A Trustworthy and Decentralized Neural Framework for Proactive Anomaly Defense in Consumer IoT

    cs.CR 2026-06 unverdicted novelty 3.0

    SwarmSense-DNN is a proposed decentralized neural framework that integrates swarm intelligence with hierarchical federated learning and graph neural networks to achieve 95.44% anomaly detection accuracy and 67% reduce...

Reference graph

Works this paper leans on

49 extracted references · 8 linked inside Pith · cited by 2 Pith papers

  1. [1]

    FedMobile: Enabling Knowledge Contribution-aware Multi-modal Federated Learning with Incomplete Modalities,

    Y . Liu, C. Wang, and X. Yuan, “FedMobile: Enabling Knowledge Contribution-aware Multi-modal Federated Learning with Incomplete Modalities,” inProc. WWW, 2025, pp. 2775–2786

    2025

  2. [2]

    Dynamic Uncertainty-Aware Multimodal Fusion for Outdoor Health Monitoring,

    Z. Fang, Z. Lin, S. Hu, Y . Tao, Y . Deng, X. Chen, and Y . Fang, “Dynamic Uncertainty-Aware Multimodal Fusion for Outdoor Health Monitoring,” arXiv preprint arXiv:2508.09085, 2025

    Pith/arXiv arXiv 2025

  3. [3]

    LLMs and IoT: A Comprehensive Survey on Large Language Models and the Internet of Things,

    F. Sarhaddi, N. T. Nguyen, A. Zuniga, P. Hui, S. Tarkoma, H. Flores, and P. Nurmi, “LLMs and IoT: A Comprehensive Survey on Large Language Models and the Internet of Things,”Authorea Preprints, 2025

    2025

  4. [4]

    Hfedmoe: Resource-Aware Heterogeneous Federated Learning with Mixture-of-Experts,

    Z. Fang, Z. Lin, S. Hu, Y . Ma, Y . Tao, Y . Deng, X. Chen, and Y . Fang, “Hfedmoe: Resource-Aware Heterogeneous Federated Learning with Mixture-of-Experts,”arXiv preprint arXiv:2601.00583, 2026

    arXiv 2026

  5. [5]

    Unleashing the Power of Continual Learning on Non-centralized Devices: A Survey,

    Y . Li, H. Wang, W. Xu, T. Xiao, H. Liu, M. Tu, Y . Wang, X. Yang, R. Zhang, S. Yuet al., “Unleashing the Power of Continual Learning on Non-centralized Devices: A Survey,”IEEE Commun. Surv. Tutor., 2025

    2025

  6. [6]

    Centralized and Federated Learning for Predictive VNF Autoscaling in Multi-domain 5G Networks and Beyond,

    T. Subramanya and R. Riggio, “Centralized and Federated Learning for Predictive VNF Autoscaling in Multi-domain 5G Networks and Beyond,” IEEE Trans. Netw. Serv. Manag., vol. 18, no. 1, pp. 63–78, 2021

    2021

  7. [7]

    Social-aware Federated Learning: Challenges and Opportunities in Collaborative Data Training,

    A.-R. Ottun, P. C. Mane, Z. Yin, S. Paul, M. Liyanage, J. Pridmore, A. Y . Ding, R. Sharma, P. Nurmi, and H. Flores, “Social-aware Federated Learning: Challenges and Opportunities in Collaborative Data Training,” IEEE Internet Comput., vol. 27, no. 2, pp. 36–44, 2022

    2022

  8. [8]

    Automated Federated Pipeline for Parameter-efficient Fine-tuning of Large Language Models,

    Z. Fang, Z. Lin, Z. Chen, X. Chen, Y . Gao, and Y . Fang, “Automated Federated Pipeline for Parameter-efficient Fine-tuning of Large Language Models,”IEEE Trans. Mobile Comput., Apr. 2025

    2025

  9. [9]

    Amount of Data Created, Consumed, and Stored 2010-2023, with Forecasts to 2028,

    “Amount of Data Created, Consumed, and Stored 2010-2023, with Forecasts to 2028,”. Statista. May. 2024. [Online]. Available: https://www.statista.com/statistics/871513/worldwide-data-created/

    2010

  10. [10]

    Optimizing Parameter Mixing Under Constrained Communications in Parallel Federated Learning,

    X. Liu, Z. Yan, Y . Zhou, D. Wu, X. Chen, and J. H. Wang, “Optimizing Parameter Mixing Under Constrained Communications in Parallel Federated Learning,”IEEE/ACM Trans. Networking, vol. 31, no. 6, pp. 2640–2652, Dec. 2023

    2023

  11. [11]

    Actions at the Edge: Jointly Optimizing the Resources in Multi-access Edge Computing,

    Y . Deng, X. Chen, G. Zhu, Y . Fang, Z. Chen, and X. Deng, “Actions at the Edge: Jointly Optimizing the Resources in Multi-access Edge Computing,”IEEE Wireless Commun., vol. 29, no. 2, pp. 192–198, Apr. 2022

    2022

  12. [12]

    Communication-efficient Learning of Deep Networks From Decentralized Data,

    B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. Arcas, “Communication-efficient Learning of Deep Networks From Decentralized Data,” inProc. AISTATS, Apr. 2017

    2017

  13. [13]

    Federated Learning: Strategies for Improving Communication Efficiency,

    J. Koneˇcn`y, H. B. McMahan, F. X. Yu, P. Richtárik, A. T. Suresh, and D. Bacon, “Federated Learning: Strategies for Improving Communication Efficiency,”arXiv preprint arXiv:1610.05492, Oct. 2016

    Pith/arXiv arXiv 2016

  14. [14]

    Accelerating Federated Learning with Model Segmentation for Edge Networks,

    M. Hu, J. Zhang, X. Wang, S. Liu, and Z. Lin, “Accelerating Federated Learning with Model Segmentation for Edge Networks,”IEEE Trans. Green Commun. Netw., Jul. 2024

    2024

  15. [15]

    Optimizing Personalized Federated Learning through Adaptive Layer-Wise Learning,

    W. Chen, C. Yang, J. Ren, Z. Li, and Z. Wang, “Optimizing Personalized Federated Learning through Adaptive Layer-Wise Learning,” inProc. IJCAI, 2024

    2024

  16. [16]

    Trust Management of Tiny Federated Learning in Internet of Unmanned Aerial Vehicles,

    J. Zheng, J. Xu, H. Du, D. Niyato, J. Kang, J. Nie, and Z. Wang, “Trust Management of Tiny Federated Learning in Internet of Unmanned Aerial Vehicles,”IEEE Internet Things J., vol. 11, no. 12, pp. 21 046–21 060, 2024

    2024

  17. [17]

    HSplitLoRA: A Heterogeneous Split Parameter- Efficient Fine-Tuning Framework for Large Language Models,

    Z. Lin, Y . Zhang, Z. Chen, Z. Fang, X. Chen, P. Vepakomma, W. Ni, J. Luo, and Y . Gao, “HSplitLoRA: A Heterogeneous Split Parameter- Efficient Fine-Tuning Framework for Large Language Models,”arXiv preprint arXiv:2505.02795, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Split learning over Wireless Networks: Parallel Design and Resource Management,

    W. Wu, M. Li, K. Qu, C. Zhou, X. Shen, W. Zhuang, X. Li, and W. Shi, “Split learning over Wireless Networks: Parallel Design and Resource Management,”IEEE J. Sel. Areas Commun., vol. 41, no. 4, pp. 1051– 1066, Feb. 2023

    2023

  19. [19]

    Gemini: A Family of Highly Capable Multimodal Models,

    G. Team, R. Anil, S. Borgeaud, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M. Dai, A. Hauth, K. Millicanet al., “Gemini: A Family of Highly Capable Multimodal Models,”arXiv preprint arXiv:2312.11805, Dec. 2023

    Pith/arXiv arXiv 2023

  20. [20]

    HASFL: Heterogeneity- aware Split Federated Learning over Edge Computing Systems,

    Z. Lin, Z. Chen, X. Chen, W. Ni, and Y . Gao, “HASFL: Heterogeneity- aware Split Federated Learning over Edge Computing Systems,”IEEE Trans. Mobile Comput., 2026

    2026

  21. [21]

    Splitfed: When Federated Learning Meets Split Learning,

    C. Thapa, P. C. M. Arachchige, S. Camtepe, and L. Sun, “Splitfed: When Federated Learning Meets Split Learning,” inProc. AAAI, Feb. 2022

    2022

  22. [22]

    Unleashing the Tiger: Inference Attacks on Split Learning,

    D. Pasquini, G. Ateniese, and M. Bernaschi, “Unleashing the Tiger: Inference Attacks on Split Learning,” inProc. CCS, Nov. 2021

    2021

  23. [23]

    Efficient Parallel Split Learning over Resource-constrained Wireless Edge Networks,

    Z. Lin, G. Zhu, Y . Deng, X. Chen, Y . Gao, K. Huang, and Y . Fang, “Efficient Parallel Split Learning over Resource-constrained Wireless Edge Networks,”IEEE Trans. Mobile Comput., 2024. 13

    2024

  24. [24]

    NSC-SL: A Bandwidth-Aware Neural Subspace Compression for Communication-Efficient Split Learning,

    Z. Fang, M. Yang, Z. Lin, Z. Lin, Z. Fang, Z. Zhang, T. Duan, D. Huang, and S. Zhu, “NSC-SL: A Bandwidth-Aware Neural Subspace Compression for Communication-Efficient Split Learning,” inProc. ICASSSP, 2026

    2026

  25. [25]

    ESL-LEO: An Efficient Split Learning Framework over LEO Satellite Networks,

    Z. Lin, Y . Zhang, Z. Chen, Z. Fang, Y . Yang, G. Zhang, H. Yang, C. Wu, X. Chen, and Y . Gao, “ESL-LEO: An Efficient Split Learning Framework over LEO Satellite Networks,” inProc. WASA, 2025

    2025

  26. [26]

    A Bargaining Game for Personalized, Energy Efficient Split Learning over Wireless Networks,

    M. Kim, A. DeRieux, and W. Saad, “A Bargaining Game for Personalized, Energy Efficient Split Learning over Wireless Networks,”arXiv preprint arXiv:2212.06107, Dec. 2022

    Pith/arXiv arXiv 2022

  27. [27]

    Splitfed Learning Without Client-side Synchronization: Analyzing Client- side Split Network Portion Size to Overall Performance,

    P. Joshi, C. Thapa, S. Camtepe, M. Hasanuzzamana, T. Scully, and H. Afli, “Splitfed Learning Without Client-side Synchronization: Analyzing Client- side Split Network Portion Size to Overall Performance,”arXiv preprint arXiv:2109.09246, Sep. 2021

    Pith/arXiv arXiv 2021

  28. [28]

    Split Learning in 6G Edge Networks,

    Z. Lin, G. Qu, X. Chen, and K. Huang, “Split Learning in 6G Edge Networks,”IEEE Wireless Commun., 2024

    2024

  29. [29]

    Towards Collaborative Fairness in Federated Learning Under Imbalanced Covariate Shift,

    T. Yu, J. Wang, H. Wang, M. Lin, H. Liu, N. S. Yee, and F. Ma, “Towards Collaborative Fairness in Federated Learning Under Imbalanced Covariate Shift,” inProc. SIGKDD, 2025, pp. 3645–3656

    2025

  30. [30]

    FedCiR: Client-invariant Representation Learning for Federated Non-IID Features,

    Z. Li, Z. Lin, J. Shao, Y . Mao, and J. Zhang, “FedCiR: Client-invariant Representation Learning for Federated Non-IID Features,”IEEE Trans. Mobile Comput., vol. 23, no. 11, pp. 10 509–10 522, 2024

    2024

  31. [31]

    FedASA: A Personalized Federated Learning with Adaptive Model Aggregation for Heterogeneous Mobile Edge Computing,

    D. Deng, X. Wu, T. Zhang, X. Tang, H. Du, J. Kang, J. Liu, and D. Niyato, “FedASA: A Personalized Federated Learning with Adaptive Model Aggregation for Heterogeneous Mobile Edge Computing,”IEEE Trans. Mobile Comput., 2024

    2024

  32. [32]

    Very Deep Convolutional Networks for Large-scale Image Recognition,

    K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-scale Image Recognition,” inProc. ICLR, 2015

    2015

  33. [33]

    Learning Multiple Layers of Features From Tiny Images,

    A. Krizhevsky, G. Hintonet al., “Learning Multiple Layers of Features From Tiny Images,”Tech. Rep., Apr. 2009

    2009

  34. [34]

    Faster Convergence on Heterogeneous Federated Edge Learning: An Adaptive Clustered Data Sharing Approach,

    G. Hu, Y . Teng, N. Wang, and Z. Han, “Faster Convergence on Heterogeneous Federated Edge Learning: An Adaptive Clustered Data Sharing Approach,”IEEE Trans. Mobile Comput., 2025

    2025

  35. [35]

    LEO-Split: A Semi-supervised Split Learning Framework over LEO Satellite Networks,

    Z. Lin, Y . Zhang, Z. Chen, Z. Fang, C. Wu, X. Chen, Y . Gao, and J. Luo, “LEO-Split: A Semi-supervised Split Learning Framework over LEO Satellite Networks,”IEEE Trans. Mobile Comput., 2025

    2025

  36. [36]

    Multi-level Personalized Federated Learning on Heterogeneous and Long-tailed Data,

    R. Zhang, Y . Chen, C. Wu, F. Wang, and B. Li, “Multi-level Personalized Federated Learning on Heterogeneous and Long-tailed Data,”IEEE Trans. Mobile Comput., vol. 23, no. 12, pp. 12 396–12 409, 2024

    2024

  37. [37]

    AOCC-FL: Federated Learning with Aligned Overlapping via Calibrated Compensation,

    H. Wang, W. Xu, Y . Fan, R. Li, and P. Zhou, “AOCC-FL: Federated Learning with Aligned Overlapping via Calibrated Compensation,” in Proc. IEEE INFOCOM, 2023

    2023

  38. [38]

    Profit Allocation for Federated Learning,

    T. Song, Y . Tong, and S. Wei, “Profit Allocation for Federated Learning,” inProc. Big Data, 2019

    2019

  39. [39]

    GOFL: An accurate and efficient federated learning framework based on gradient optimization in Heterogeneous IoT systems,

    Z. Lian, J. Cao, Z. Zhu, X. Zhou, and W. Liu, “GOFL: An accurate and efficient federated learning framework based on gradient optimization in Heterogeneous IoT systems,”IEEE Internet Things J., vol. 11, no. 7, pp. 12 459–12 474, 2023

    2023

  40. [40]

    NVIDIA Jetson Embedded Systems Developer Kits and Mod- ules

    (2024) “NVIDIA Jetson Embedded Systems Developer Kits and Mod- ules". Available: https://www.nvidia.cn/autonomous-machines/embedded- systems/?section=jetsonTX2

    2024

  41. [41]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gellyet al., “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” inProc. ICLR, 2021

    2021

  42. [42]

    Split Learning for Health: Distributed Deep Learning without Sharing Raw Patient Data,

    P. Vepakomma, O. Gupta, T. Swedish, and R. Raskar, “Split Learning for Health: Distributed Deep Learning without Sharing Raw Patient Data,” arXiv preprint arXiv:1812.00564, Dec. 2018

    Pith/arXiv arXiv 2018

  43. [43]

    Server-side Local Gradient Averaging and Learning Rate Acceleration for Scalable Split Learning,

    S. Pal, M. Uniyal, J. Park, P. Vepakomma, R. Raskar, M. Bennis, M. Jeon, and J. Choi, “Server-side Local Gradient Averaging and Learning Rate Acceleration for Scalable Split Learning,”arXiv preprint arXiv:2112.05929, Dec. 2021

    Pith/arXiv arXiv 2021

  44. [44]

    Adaptsfl: Adaptive Split Federated Learning in Resource-constrained Edge Networks,

    Z. Lin, G. Qu, W. Wei, X. Chen, and K. K. Leung, “Adaptsfl: Adaptive Split Federated Learning in Resource-constrained Edge Networks,”IEEE Trans. Networking, Jul. 2025

    2025

  45. [45]

    Hierarchical Split Federated Learning: Convergence Analysis and System Optimization,

    Z. Lin, W. Wei, Z. Chen, C.-T. Lam, X. Chen, Y . Gao, and J. Luo, “Hierarchical Split Federated Learning: Convergence Analysis and System Optimization,”IEEE Trans. Mobile Comput., Dec. 2024

    2024

  46. [46]

    FedAW A: Adaptive Optimization of Aggregation Weights in Federated Learning Using Client Vectors,

    C. Shi, H. Zhao, B. Zhang, M. Zhou, D. Guo, and Y . Chang, “FedAW A: Adaptive Optimization of Aggregation Weights in Federated Learning Using Client Vectors,” inProc. CVPR, 2025

    2025

  47. [47]

    Scaffold: Stochastic Controlled Averaging for Federated Learning,

    S. P. Karimireddy, S. Kale, M. Mohri, S. Reddi, S. Stich, and A. T. Suresh, “Scaffold: Stochastic Controlled Averaging for Federated Learning,” in Proc. ICML, Apr. 2020

    2020

  48. [48]

    Fedcache: A Knowledge Cache-driven Federated Learning Architecture for Personalized Edge Intelligence,

    Z. Wu, S. Sun, Y . Wang, M. Liu, K. Xu, W. Wang, X. Jiang, B. Gao, and J. Lu, “Fedcache: A Knowledge Cache-driven Federated Learning Architecture for Personalized Edge Intelligence,”IEEE Trans. Mobile Comput., vol. 23, no. 10, pp. 9368–9382, 2024

    2024

  49. [49]

    Towards Personalized Federated Learning via Heterogeneous Model Reassembly,

    J. Wang, X. Yang, S. Cui, L. Che, L. Lyu, D. D. Xu, and F. Ma, “Towards Personalized Federated Learning via Heterogeneous Model Reassembly,” Proc. NIPS, 2023

    2023