pith. sign in

arxiv: 2606.09869 · v1 · pith:PRWXRPZJnew · submitted 2026-06-01 · 💻 cs.LG · cs.AI· cs.CR

QSplitFL: Capability Aware Deep Q-Learning for Optimal Split Point Selection in Split Federated Learning

Nazmus Shakib Shadin , Xinyue Zhang , Jingyi Wang , Miao Pan This is my paper

Pith reviewed 2026-06-28 15:37 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CR
keywords split federated learningdeep Q-networksplit point selectionheterogeneous devicesresource-aware learningfederated learningsplit learning
0
0 comments X

The pith

QSplitFL's deep Q-network selects split points using only client hardware metrics to speed convergence in split federated learning on heterogeneous devices.

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

The paper presents QSplitFL, a framework that uses deep Q-learning to choose where to split neural network models in split federated learning setups. It addresses the problem of fixed split points overloading weaker devices in groups with varied hardware. By representing device state through simple metrics like CPU use, memory, battery, and latency, the system learns good split decisions without needing full model details. Experiments across image datasets and common network architectures show faster convergence and better accuracy than prior methods. This matters because it makes privacy-preserving collaborative training practical on real-world mixed-device networks.

Core claim

QSplitFL is a capability-aware Deep Q-Network framework for optimal split point selection in Split learning based Federated Learning. It uses a lightweight state representation from client hardware metrics and a decayed loss-drop reward function with a committee-based DQN architecture. Extensive experiments demonstrate better convergence and higher accuracy compared to existing methods while adapting to heterogeneous device resources on multiple datasets and architectures.

What carries the argument

The Deep Q-Network that takes hardware metrics as state input and outputs split point actions, trained with decayed loss-drop rewards and committee voting.

If this is right

  • Split points adapt dynamically to device capabilities, reducing overload on weak clients.
  • Convergence speeds up and final accuracy increases on standard benchmarks.
  • The approach works across CNNs to modern architectures like ConvNeXt without model weight inputs.
  • Reward design prioritizes early training progress over later stages.

Where Pith is reading between the lines

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

  • The method may extend to other split decisions in distributed training beyond federated settings.
  • Hardware-only states could enable privacy benefits by avoiding sharing model internals.
  • Testing on real mobile hardware rather than simulated metrics would strengthen the results.

Load-bearing premise

Client hardware metrics alone provide sufficient information for the DQN to learn optimal split points without access to model weights or internal states.

What would settle it

Running the trained QSplitFL agent on a new device fleet whose hardware distributions differ markedly from the training set and observing that the chosen splits yield slower convergence or lower accuracy than a fixed middle-layer split would falsify the claim.

Figures

Figures reproduced from arXiv: 2606.09869 by Jingyi Wang, Miao Pan, Nazmus Shakib Shadin, Xinyue Zhang.

Figure 1
Figure 1. Figure 1: QSplitFL Workflow Architecture. (1) Client-Side: Clients has hard￾ware metrics, receive split layer ℓ, and run forward propagation through layers 1 to ℓ, which produces smashed data (Ak). (2) Server-Side: The server completes training through layers ℓ + 1 to L and returns gradients to clients. (3) Aggre￾gation: Client updates are aggregated via FedAvg; reward (rt) is computed on the based of the loss funct… view at source ↗
Figure 2
Figure 2. Figure 2: High-Level Overview of QSplitFL Framework Operation in Each Training Round. The diagram illustrates the complete workflow: (1) capabil￾ity metric collection and state construction, (2) committee-based split point selection, (3) SFL execution with smashed data transmission, and (4) federated aggregation with reward computation. With this per-round view in place, we now formalize each component, begin￾ning w… view at source ↗
Figure 3
Figure 3. Figure 3: Accuracy Convergence Analysis (100 Rounds). Comparison of ResNet50, MobileNetV4, and ConvNeXt across MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100 with varying client counts. CNN results are detailed in Appendix A.9. that the agent focuses primarily on the model architecture rather than the num￾ber of participating clients when making split decisions. The consistency of these choices across datasets also i… view at source ↗
Figure 4
Figure 4. Figure 4: Dynamic Split Point Selection (100 Rounds). Adaptive split point selection for ResNet50, MobileNetV4, and ConvNeXt. The agent consistently selects optimal split layers (30–59 depending on architecture) to balance com￾putation and communication. regardless of the model architectures, number of clients, and training rounds, which ensures that deeper and more modern architectures benefit the most from QSplitF… view at source ↗
Figure 5
Figure 5. Figure 5: Challenges in Existing Split Federated Learning vs Our Pro￾posed Solution. The left side illustrates the key limitations of current state￾of-the-art approaches, including static split points, inability to handle device heterogeneity. The right side presents QSplitFL’s capability-aware SFL based reinforcement learning solution that dynamically adapts split points based on current client resource metrics. A.… view at source ↗
Figure 6
Figure 6. Figure 6: Committee-Based Reward Hacking Prevention Mechanism. The architecture employs M DQN members (typically M = 3 or M = 5, always odd) which shares a common encoder but maintains independent decision heads. Each member proposes its preferred split action, and the final decision is made through majority voting. This ensemble approach mitigates reward hacking by ensuring decisions reflect consensus across divers… view at source ↗
Figure 7
Figure 7. Figure 7: Comprehensive Model Architecture Performance Comparison Across All [PITH_FULL_IMAGE:figures/full_fig_p031_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: CNN Architecture Split Point Selection Analysis. [PITH_FULL_IMAGE:figures/full_fig_p032_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: CNN Accuracy Convergence (100 Rounds). Stable convergence on simple tasks contrasts with limited performance on CIFAR-100 due to shallow depth [PITH_FULL_IMAGE:figures/full_fig_p033_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: MNIST Dataset Accuracy Convergence (10, 20, 50 Rounds). Comprehensive accuracy convergence analysis across four model architectures (CNN, ResNet50, MobileNetV4, ConvNeXt) and varying training durations. Each row corresponds to one architecture, with columns showing 10, 20, and 50 training rounds. All architectures achieve rapid convergence on this baseline dataset, reaching >95% accuracy within 10 rounds … view at source ↗
Figure 11
Figure 11. Figure 11: Fashion-MNIST Accuracy Convergence (10, 20, 50 Rounds). Accuracy convergence comparison across architectures and training rounds for the Fashion-MNIST dataset. The figure shows CNN (row 1), ResNet50 (row 2), MobileNetV4 (row 3), and ConvNeXt (row 4) performance at 10, 20, and 50 rounds. This dataset is more structurally complex than MNIST: ConvNeXt consistently achieves the highest accuracy (∼94% at 50 ro… view at source ↗
Figure 12
Figure 12. Figure 12: CIFAR-10 Accuracy Convergence (10, 20, 50 Rounds). Con￾vergence trends for CIFAR-10 dataset shows performance differentiation among different architectures. The figure illustrates that 10-round training is insufficient for effective feature learning, particularly for shallow CNN. Extended training to 50 rounds enables deep architectures (ResNet50, ConvNeXt) to achieve substan￾tially higher accuracy (∼86%)… view at source ↗
Figure 13
Figure 13. Figure 13: CIFAR-100 Accuracy Convergence (10, 20, 50 Rounds). Per￾formance analysis on the most challenging fine grained classification task with 100 classes. The 4×3 grid (architectures × rounds) reveals architectural impact: shallow CNN struggles to capture fine grained features even at 50 rounds (∼62% accuracy), while ConvNeXt reaches ∼68%. The progressive improvement from 10 to 50 rounds demonstrates that QSpli… view at source ↗
read the original abstract

Federated Learning (FL) combined with Split Learning (SL) is a privacy preserving paradigm that enables training deep neural networks (DNNs) on resource constrained devices while reducing overall training cost. However, determining the optimal split point, meaning the layer where the model is divided still remains a critical challenge, especially when clients have heterogeneous hardware capabilities. Fixed split points can overload weak devices and increase the communication and server load, which slows convergence and reduces stability. This paper introduces QSplitFL, a novel capability-aware Deep Q-Network (DQN) framework for optimal split point selection in Split learning based Federated Learning (SFL) environments. Unlike existing approaches that rely on high-dimensional model weight representations, QSplitFL employs a lightweight state representation derived directly from client hardware metrics, including CPU utilization, memory, battery level, and network latency. The proposed framework incorporates a decayed loss-drop reward function that prioritizes early convergence, and a committee-based DQN architecture with majority voting to mitigate reward hacking. Extensive experiments on MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100 datasets using CNN, ResNet50, MobileNetV4, and ConvNeXt architectures demonstrate that our approach achieves better convergence and higher accuracy compared to existing methods, while effectively adapting to heterogeneous device resources. The source code is publicly available at https://github.com/AIPO-Lab/QSplitFL.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The paper proposes QSplitFL, a Deep Q-Network framework for optimal split-point selection in Split Federated Learning. It replaces high-dimensional model-weight states with a lightweight representation using only client hardware metrics (CPU utilization, memory, battery level, network latency). A decayed loss-drop reward prioritizes early convergence and a committee-based DQN with majority voting mitigates reward hacking. Experiments across MNIST, Fashion-MNIST, CIFAR-10, CIFAR-100 and CNN/ResNet50/MobileNetV4/ConvNeXt architectures are asserted to yield faster convergence and higher accuracy than existing methods while adapting to device heterogeneity; source code is released publicly.

Significance. If the empirical claims hold with verifiable metrics, the work would be significant for practical SFL on heterogeneous edge devices by demonstrating that low-dimensional hardware states suffice for effective RL-based split selection, reducing communication and compute overhead. Public code availability is a clear strength supporting reproducibility. The result would be less impactful if the hardware-only state proves architecture-specific, limiting transfer to new models.

major comments (3)
  1. [Abstract] Abstract: the central claim that QSplitFL 'achieves better convergence and higher accuracy' rests on 'extensive experiments' yet no quantitative tables, accuracy values, convergence curves, baseline definitions, or statistical tests are referenced; this absence is load-bearing because the empirical superiority cannot be assessed.
  2. [Method] Method (state representation): the DQN state is defined solely from hardware metrics with no layer descriptors or model features. Because per-layer compute costs vary sharply across architectures (early conv layers vs. later FC layers in ResNet50 vs. ConvNeXt), the learned mapping from hardware state to split point is likely tied to the four training architectures; this directly undermines the claim of effective adaptation under heterogeneity for unseen models.
  3. [Method] Reward design: the decayed loss-drop reward and committee voting mechanism are introduced to favor early convergence and prevent hacking, yet both contain free parameters (decay schedule, voting threshold) that can be tuned post-hoc to observed training curves, introducing circularity between the modeling choices and the reported performance.
minor comments (1)
  1. [Abstract] Abstract: nomenclature 'MobileNetV4' is non-standard; confirm whether this refers to MobileNetV2, V3, or a custom variant.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment point by point below, with clarifications and planned revisions where appropriate to improve the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that QSplitFL 'achieves better convergence and higher accuracy' rests on 'extensive experiments' yet no quantitative tables, accuracy values, convergence curves, baseline definitions, or statistical tests are referenced; this absence is load-bearing because the empirical superiority cannot be assessed.

    Authors: We agree that the abstract would benefit from explicit quantitative support. In the revision we will insert concise references to key results (e.g., accuracy gains and convergence epochs versus baselines) together with pointers to the corresponding tables and figures. revision: yes

  2. Referee: [Method] Method (state representation): the DQN state is defined solely from hardware metrics with no layer descriptors or model features. Because per-layer compute costs vary sharply across architectures (early conv layers vs. later FC layers in ResNet50 vs. ConvNeXt), the learned mapping from hardware state to split point is likely tied to the four training architectures; this directly undermines the claim of effective adaptation under heterogeneity for unseen models.

    Authors: Hardware metrics were selected precisely because they are architecture-agnostic and reflect device capability. Experiments show the policy adapts across the four evaluated architectures under heterogeneous conditions. We will add an explicit limitations paragraph noting that generalization to entirely new architectures would require retraining and is not claimed in the current work. revision: partial

  3. Referee: [Method] Reward design: the decayed loss-drop reward and committee voting mechanism are introduced to favor early convergence and prevent hacking, yet both contain free parameters (decay schedule, voting threshold) that can be tuned post-hoc to observed training curves, introducing circularity between the modeling choices and the reported performance.

    Authors: The decay schedule and voting threshold were fixed after preliminary validation runs conducted before the final experiments. We will report the exact values used and include a sensitivity study demonstrating that performance remains stable under modest perturbations of these parameters. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical RL method with independent experimental validation

full rationale

The paper proposes a DQN-based framework for split-point selection whose state uses only client hardware metrics and whose reward is a decayed loss-drop function. These are explicit design choices in the RL formulation, not a claimed first-principles derivation. Performance claims rest on direct empirical comparisons across four datasets and four architectures; no equation or theorem is shown to reduce the reported accuracy or convergence gains to the reward definition or state representation by algebraic identity. No self-citation chain is invoked to justify uniqueness or load-bearing assumptions. The method is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 1 invented entities

The framework depends on several RL hyperparameters and a domain assumption that hardware metrics alone suffice; the committee architecture is an invented component introduced to address reward hacking.

free parameters (2)
  • decayed loss-drop reward decay schedule
    Parameters controlling how quickly the reward decays are chosen or fitted to produce the reported early-convergence behavior.
  • DQN training hyperparameters
    Learning rate, exploration schedule, network sizes, and batch sizes are standard free parameters that must be set for the reported results.
axioms (1)
  • domain assumption Client hardware metrics are a sufficient state representation for optimal split-point decisions
    The method is built on the premise that CPU, memory, battery, and latency readings alone allow the DQN to select good splits without model-specific features.
invented entities (1)
  • Committee-based DQN with majority voting no independent evidence
    purpose: Mitigate reward hacking in the reinforcement learning loop
    This multi-network voting mechanism is introduced by the paper to stabilize training and is not an external benchmark.

pith-pipeline@v0.9.1-grok · 5797 in / 1582 out tokens · 36327 ms · 2026-06-28T15:37:16.678296+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

37 extracted references · 12 canonical work pages · 3 internal anchors

  1. [1]

    In: Proceedings of the 2016 ACM SIGSAC conference on computer and communications security

    Abadi, M., Chu, A., Goodfellow, I., McMahan, H.B., Mironov, I., Talwar, K., Zhang, L.: Deep learning with differential privacy. In: Proceedings of the 2016 ACM SIGSAC conference on computer and communications security. pp. 308–318 (2016)

    2016

  2. [2]

    Sensors22(2), 450 (2022) QSplitFL: Deep Q-Learning for Split Point Selection in SFL 17

    Abreha, H.G., Hayajneh, M., Serhani, M.A.: Federated learning in edge computing: a systematic survey. Sensors22(2), 450 (2022) QSplitFL: Deep Q-Learning for Split Point Selection in SFL 17

    2022

  3. [3]

    Journal of Network and Computer Applications236, 104105 (2025)

    Arafeh, M., Wazzeh, M., Sami, H., Ould-Slimane, H., Talhi, C., Mourad, A., Otrok, H.: Efficient privacy-preserving ml for iot: Cluster-based split federated learning scheme for non-iid data. Journal of Network and Computer Applications236, 104105 (2025)

    2025

  4. [4]

    Flower: A Friendly Federated Learning Research Framework

    Beutel, D.J., Topal, T., Mathur, A., Qiu, X., Fernandez-Marques, J., Gao, Y., Sani, L., Li, K.H., Parcollet, T., de Gusmão, P.P.B., et al.: Flower: A friendly federated learning research framework. arXiv preprint arXiv:2007.14390 (2020)

    work page internal anchor Pith review Pith/arXiv arXiv 2007

  5. [5]

    In: 2024 19th International Symposium on Wireless Communication Systems (ISWCS)

    Binucci, F., Merluzzi, M., Banelli, P., Strinati, E.C., Di Lorenzo, P.: Enabling edge artificial intelligence via goal-oriented deep neural network splitting. In: 2024 19th International Symposium on Wireless Communication Systems (ISWCS). pp. 1–6. IEEE (2024)

    2024

  6. [6]

    Sensors24(1), 88 (2023)

    Chen, H., Chen, X., Peng, L., Bai, Y.: Personalized fair split learning for resource- constrained internet of things. Sensors24(1), 88 (2023)

    2023

  7. [7]

    IEEE Internet of Things Journal12(15), 30460–30474 (2025).https://doi.org/10.1109/JIOT.2025.3572393

    Chen, X., Li, J., Fan, D., Chakrabarti, C.: Heterosfl: Split federated learning with heterogeneous clients and non-iid data. IEEE Internet of Things Journal12(15), 30460–30474 (2025).https://doi.org/10.1109/JIOT.2025.3572393

    work page doi:10.1109/jiot.2025.3572393 2025

  8. [8]

    arXiv preprint arXiv:2412.15536 (2024)

    Dachille, J., Huang, C., Liu, X.: The impact of cut layer selection in split federated learning. arXiv preprint arXiv:2412.15536 (2024)

    work page arXiv 2024

  9. [9]

    IEEE Transactions on Mobile Computing (2025)

    Fan, W., Chen, P., Chun, X., Liu, Y.: Madrl-based model partitioning, aggregation control, and resource allocation for cloud-edge-device collaborative split federated learning. IEEE Transactions on Mobile Computing (2025)

    2025

  10. [10]

    Journal of Network and Computer Applications116, 1–8 (2018)

    Gupta, O., Raskar, R.: Distributed learning of deep neural network over multiple agents. Journal of Network and Computer Applications116, 1–8 (2018)

    2018

  11. [11]

    International Journal of Computing and Digital Sys- tems11(1), 541–552 (2022)

    Hariharan, N., Paavai, A.G.: A brief study of deep reinforcement learning with epsilon-greedy exploration. International Journal of Computing and Digital Sys- tems11(1), 541–552 (2022)

    2022

  12. [12]

    IEEE Access13, 46312 (2025)

    Hukkeri,G.S.,Goudar,R.,Dhananjaya,G.,Rathod,V.N.:Acomprehensivesurvey on split-fed learning: methods, innovations, and future directions. IEEE Access13, 46312 (2025)

    2025

  13. [13]

    IEEE Com- munications Surveys & Tutorials23(3), 1759–1799 (2021)

    Khan, L.U., Saad, W., Han, Z., Hossain, E., Hong, C.S.: Federated learning for internet of things: Recent advances, taxonomy, and open challenges. IEEE Com- munications Surveys & Tutorials23(3), 1759–1799 (2021)

    2021

  14. [14]

    In: 2019 Asia-Pacific signal and information processing association annual summit and conference (APSIPA ASC)

    Lee, Y.L., Qin, D.: A survey on applications of deep reinforcement learning in resource management for 5g heterogeneous networks. In: 2019 Asia-Pacific signal and information processing association annual summit and conference (APSIPA ASC). pp. 1856–1862. IEEE (2019)

    2019

  15. [15]

    In: 2024 IEEE International Conference on E-health Networking, Application & Services (HealthCom)

    Li, J., Yu, S.: Integrity verifiable privacy-preserving federated learning for healthcare-iot. In: 2024 IEEE International Conference on E-health Networking, Application & Services (HealthCom). pp. 1–6. IEEE (2024)

    2024

  16. [16]

    arXiv preprint arXiv:2102.08504 (2021)

    Li, O., Sun, J., Yang, X., Gao, W., Zhang, H., Xie, J., Smith, V., Wang, C.: Label leakage and protection in two-party split learning. arXiv preprint arXiv:2102.08504 (2021)

    work page arXiv 2021

  17. [17]

    Proceedings of Machine learning and sys- tems2, 429–450 (2020)

    Li, T., Sahu, A.K., Zaheer, M., Sanjabi, M., Talwalkar, A., Smith, V.: Federated optimization in heterogeneous networks. Proceedings of Machine learning and sys- tems2, 429–450 (2020)

    2020

  18. [18]

    arXiv preprint arXiv:1905.10497 (2019)

    Li, T., Sanjabi, M., Beirami, A., Smith, V.: Fair resource allocation in federated learning. arXiv preprint arXiv:1905.10497 (2019)

    work page arXiv 1905

  19. [19]

    arXiv preprint arXiv:1907.02189 (2019) 18 Nazmus Shakib Shadin, Xinyue Zhang, Jingyi Wang, and Miao Pan

    Li, X., Huang, K., Yang, W., Wang, S., Zhang, Z.: On the convergence of fedavg on non-iid data. arXiv preprint arXiv:1907.02189 (2019) 18 Nazmus Shakib Shadin, Xinyue Zhang, Jingyi Wang, and Miao Pan

    work page arXiv 1907

  20. [20]

    arXiv preprint arXiv:2501.01078 (2025)

    Liang, Y., Chen, Q., Zhu, G., Awan, M.K., Jiang, H.: Communication-and- computation efficient split federated learning: Gradient aggregation and resource management. arXiv preprint arXiv:2501.01078 (2025)

    work page arXiv 2025

  21. [21]

    IEEE Transactions on Wireless Communications 22(4), 2650–2665 (2022)

    Liu, X., Deng, Y., Mahmoodi, T.: Wireless distributed learning: A new hybrid split and federated learning approach. IEEE Transactions on Wireless Communications 22(4), 2650–2665 (2022)

    2022

  22. [22]

    Nixon, J., Lakshminarayanan, B., Tran, D.: Why are bootstrapped deep ensembles not better? In: ”I Can’t Believe It’s Not Better!” NeurIPS 2020 workshop (2020)

    2020

  23. [23]

    Advances in neural information processing systems29(2016)

    Osband, I., Blundell, C., Pritzel, A., Van Roy, B.: Deep exploration via boot- strapped dqn. Advances in neural information processing systems29(2016)

    2016

  24. [24]

    Computer Networks218, 109380 (2022)

    Samikwa, E., Di Maio, A., Braun, T.: Ares: Adaptive resource-aware split learning for internet of things. Computer Networks218, 109380 (2022)

    2022

  25. [25]

    Prioritized Experience Replay

    Schaul, T., Quan, J., Antonoglou, I., Silver, D.: Prioritized experience replay. arXiv preprint arXiv:1511.05952 (2015)

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  26. [26]

    arXiv preprint arXiv:1909.09145 (2019)

    Singh, A., Vepakomma, P., Gupta, O., Raskar, R.: Detailed comparison of com- munication efficiency of split learning and federated learning. arXiv preprint arXiv:1909.09145 (2019)

    work page arXiv 1909

  27. [27]

    In: Proceedings of the AAAI conference on artificial intelligence

    Thapa, C., Arachchige, P.C.M., Camtepe, S., Sun, L.: Splitfed: When federated learning meets split learning. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 8485–8493 (2022)

    2022

  28. [28]

    arXiv preprint arXiv:2508.08339 (2025)

    Tran, D.T., Ha, N.B., Nguyen, V.D., Wong, K.S.: Sherl-fl: When representation learning meets split learning in hierarchical federated learning. arXiv preprint arXiv:2508.08339 (2025)

    work page arXiv 2025

  29. [29]

    Split learning for health: Distributed deep learning without sharing raw patient data

    Vepakomma, P., Gupta, O., Swedish, T., Raskar, R.: Split learning for health: Distributed deep learning without sharing raw patient data. arXiv preprint arXiv:1812.00564 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  30. [30]

    In: IEEE INFOCOM 2020-IEEE conference on computer communications

    Wang, H., Kaplan, Z., Niu, D., Li, B.: Optimizing federated learning on non-iid data with reinforcement learning. In: IEEE INFOCOM 2020-IEEE conference on computer communications. pp. 1698–1707. IEEE (2020)

    2020

  31. [31]

    IEEE Internet of Things Journal9(21), 20889–20901 (2022)

    Wu, D., Ullah, R., Harvey, P., Kilpatrick, P., Spence, I., Varghese, B.: Fedadapt: Adaptive offloading for iot devices in federated learning. IEEE Internet of Things Journal9(21), 20889–20901 (2022)

    2022

  32. [32]

    arXiv preprint arXiv:2407.03038 (2024)

    Wu, F., Liu, X., Wang, H., Wang, X., Su, L., Gao, J.: Towards federated rlhf with aggregated client preference for llms. arXiv preprint arXiv:2407.03038 (2024)

    work page arXiv 2024

  33. [33]

    IEEE Wireless Communications31(3), 177–184 (2023)

    Wu, M., Cheng, G., Li, P., Yu, R., Wu, Y., Pan, M., Lu, R.: Split learning with differential privacy for integrated terrestrial and non-terrestrial networks. IEEE Wireless Communications31(3), 177–184 (2023)

    2023

  34. [34]

    Electronics12(11), 2478 (2023)

    Yan, S., Zhang, P., Huang, S., Wang, J., Sun, H., Zhang, Y., Tolba, A.: Node selection algorithm for federated learning based on deep reinforcement learning for edge computing in iot. Electronics12(11), 2478 (2023)

    2023

  35. [35]

    IEEE Transactions on Intelligent Trans- portation Systems (2025)

    Yu, L., Chang, Z., Jia, Y., Min, G.: Model partition and resource allocation for split learning in vehicular edge networks. IEEE Transactions on Intelligent Trans- portation Systems (2025)

    2025

  36. [36]

    IEEE Transactions on Network Science and Engineering10(3), 1339–1351 (2022)

    Yuan, X., Zhang, Z., Feng, C., Cui, Y., Garg, S., Kaddoum, G., Yu, K.: A dqn- based frame aggregation and task offloading approach for edge-enabled iomt. IEEE Transactions on Network Science and Engineering10(3), 1339–1351 (2022)

    2022

  37. [37]

    F i x e d

    Zhu, G., Deng, Y., Chen, X., Zhang, H., Fang, Y., Wong, T.F.: Esfl: Efficient split federated learning over resource-constrained heterogeneous wireless devices. IEEE Internet of Things Journal11(16), 27153–27166 (2024) QSplitFL: Deep Q-Learning for Split Point Selection in SFL 19 A Appendix This appendix provides detailed supplementary materials that corr...

    2024