REVIEW 2 major objections 5 minor 1 cited by
How Can Incentives and Cut Layer Selection Influence Data Contribution in Split Federated Learning?
T0 review · 2 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash
Pith's one-line read The paper claims that split federated learning is best understood as a Stackelberg game in which the model owner's choice of incentive and cut layer steers clients' data contributions, and it derives a closed-form Nash equilibrium for…
desk verdict A genuine new application of Stackelberg games to split federated learning, but the headline closed-form equilibrium is only an interior solution and the paper overstates its generality. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The engine of the analysis is the clients' utility function under proportional incentive sharing: $$U_n = \psi_n R \frac{d_n}{\sum_{l\in\mathcal N} d_l} - d_n H_n - I_n + S,$$ where a client receives an incentive proportional to its share of total contributed data, pays an energy cost linear in its own contribution, and has a fixed offset $S$. Taking first-order conditions and solving the resulting fixed point gives the closed-form Nash equilibrium of Theorem 4. The owner's utility is $$U_{MO} = \tau_1 \ln\left(1 + \frac{\sum_n d_n}{d_{\mathrm{req}}}\right) + \tau_2 \frac{f_{\mathrm{FLOP}}(L_c)}{w_{\mathrm{FLOP}}} - R,$$ which is strictly concave in $R$ after substituting the equilibrium $d_n^* = X_n(L_c)R$; exhaustive search over $L_c$ then yields the Stackelberg equilibrium.
What would settle it
Run the mechanism with a client whose dataset size $D_n$ is smaller than the value of $d_n^*$ predicted by the closed form; whenever any such bound is active, the predicted aggregate contribution and owner utility will not match the actual outcome. A direct computational test is to solve the constrained best-response dynamics and compare the result with (20) across a range of $R$ where one client's bound binds.
Extended reading notes
Core claim
For a fixed incentive $R$ and cut layer $L_c$, the clients' subgame has a unique Nash equilibrium in which client $n$ contributes $$d_n^* = \frac{(N-1)R}{\sum_{l\in\mathcal N} H_l/\psi_l}\left(1 - \frac{H_n (N-1)}{\psi_n \sum_{l\in\mathcal N} H_l/\psi_l}\right),$$ where $H_n$ is the marginal energy cost of contributing data and $\psi_n$ is the client's valuation of incentives. Because this equilibrium contribution is linear in $R$ for fixed $L_c$, the model owner's utility becomes strictly concave in $R$, so the owner can find its optimal incentive with a convex solver and search over the relatively small cut-layer space; concatenating that response with the clients' equilibrium yields a Stackelberg equilibrium. The paper claims that this equilibrium is a desirable operating point for both the model owner and the clients in competitive settings, and that it balances utility, accuracy, and privacy better than simply maximizing or minimizing the incentive.
Load-bearing premise
The closed-form equilibrium and the owner's optimization assume that, at the chosen incentive and cut layer, every client's best response lies strictly inside its dataset bounds, so no client is forced to contribute zero or hits its maximum dataset size.
Editorial extensions
If this is right
- Larger incentive $R$ increases every client's equilibrium contribution, but the exact slope depends on the client's energy cost $H_n$ and incentive valuation $\psi_n$; the closed form gives this slope explicitly.
- Clients with higher CPU capability, meaning lower marginal energy cost $H_n$, contribute less at the same incentive, because they value the trade-off between earning rewards and conserving energy differently.
- Raising the cut layer $L_c$ increases client-side model complexity and communication cost, which reduces equilibrium data contributions, while also making reconstruction attacks harder and thereby improving privacy.
- The owner's optimal incentive balances data satisfaction and workload offload against payment: when the owner weights server load reduction heavily relative to data satisfaction, it chooses the maximal cut layer and a smaller incentive.
- The price of anarchy of the clients' noncooperative game decreases as the number of clients or the cut layer increases, meaning the Nash equilibrium becomes closer to the centralized optimum when competition is stronger.
Reading between the lines
- The linearity $d_n^* = X_n(L_c)R$ implies that in the interior regime, every client's equilibrium contribution is exactly proportional to the total incentive, so the incentive elasticity of data contribution is 1; this is a testable prediction that could be checked in a controlled experiment.
- The same backward-induction template would extend to other proportional reward-sharing rules, such as sharing incentives by model quality or by Shapley contributions, although the closed form would change and may not remain linear.
- Because differential privacy noise degrades accuracy more at shallow cut layers than at deep ones, the paper's static game could be extended to let the owner choose $L_c$ partly to reduce the accuracy cost of the clients' chosen DP noise level.
- The interior-equilibrium assumption suggests a boundary regime the paper does not analyze: when a client's dataset size $D_n$ is smaller than the predicted $d_n^*$, the closed form fails and the owner's optimization would need to account for capacity-constrained clients explicitly.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper models strategic interaction in split federated learning (SFL) as a single-leader multi-follower Stackelberg game. The leader (SFL model owner) chooses a cut layer L_c and a total incentive R, while the followers (clients) choose how many data samples d_n to contribute. The clients' utility is a proportional share of R minus energy cost, and the owner's utility balances data satisfaction, load reduction, and incentive expenditure. The central theoretical claims are: existence and uniqueness of the client-level Nash equilibrium with a closed-form expression (Theorem 4, Eq. (20)); existence of an optimal owner strategy (Theorem 5); and existence of a Stackelberg equilibrium (Theorem 6). The paper also contains classification experiments on CIFAR-10, Fashion-MNIST, and CIFAR-100 to illustrate accuracy gains from higher incentives.
Significance. If the closed-form equilibrium and the Stackelberg equilibrium results were fully correct, the paper would offer a simple, tractable tool for incentive and cut-layer design in SFL. The utility models are stylized but reasonable, and the explicit formula could be useful for system planning. The paper is one of the first to treat SFL as a competitive game, which is a meaningful conceptual step. However, the analysis relies on an interiority assumption that is not stated or verified, and the uniqueness proof is conditional in a way that undercuts the theorems as stated. The experimental section is illustrative rather than an independent validation of the equilibrium predictions.
major comments (2)
- [Section IV.A, Theorem 4 and Eq. (20)] The closed-form Nash equilibrium in Eq. (20) is derived only from the interior branch of the best-response function. The actual client problem (13) includes the constraints 0 ≤ d_n ≤ D_n, and the paper's own best-response expression (19) is piecewise, with zero and D_n branches. The proof of Theorem 4 uses Eq. (21), which is obtained by substituting the middle branch of (19), and this equation holds only if every client's best response is strictly inside (0, D_n). When boundary conditions bind, Eq. (20) is not the Nash equilibrium. A concrete failure: for N=3 with H_1/ψ_1 = 100 and H_2/ψ_2 = H_3/ψ_3 = 1, Eq. (20) gives d_1^* < 0, while the constrained best-response logic in (19) yields d_1^* = 0 with positive contributions from clients 2 and 3. Similarly, for large R, Eq. (20) can exceed D_n. Stage 2 then substitutes d_n^* = X_n(L_c)R in Eq. (26) and Algorithm 1 maximizes (25) over this expression, so the claimed Stackelberg equilibrium is computed against a response function that is not the true client response when any bound is active. The theorem should state explicit conditions sufficient for 0 < d_n^* < D_n for all n, or the analysis must be extended to the constrained case via a KKT treatment.
- [Section IV.A, Theorems 2 and 3] The uniqueness proof is conditional, but the theorem is stated unconditionally. Theorem 2 claims that the best response is a standard function for 'sufficiently large R,' but the proof requires extra inequalities, e.g., ψ_n R / H_n > 4∑_{l≠n} d_l, to establish positivity and monotonicity. Moreover, the best-response map in Eq. (19) has a zero branch, so it is not positive on the whole domain as required by Definition 3. Thus the proof of Theorem 2 only establishes the standard-function property under additional assumptions that are not incorporated into Theorem 3. Theorem 3 then asserts a unique Nash equilibrium without any qualifier, and Theorem 6 uses this uniqueness to conclude the existence of a Stackelberg equilibrium. The theorems need to either include the sufficient conditions explicitly or be replaced with a different, globally valid uniqueness argument.
minor comments (5)
- [Section IV.A, Eq. (19)] The middle branch of Eq. (19) is labeled 'p18q' rather than '(18)'.
- [Author affiliation] The affiliation line contains the typo 'Rep. ok Korea' instead of 'Rep. of Korea'.
- [Reference [28]] Reference [28] misspells 'Toronto' as 'Tronto'.
- [Section V, Fig. 6] The classification experiments in Fig. 6 allocate training samples to clients using the model's own equilibrium formula (20) for each incentive value R. These experiments therefore demonstrate that more training data improves accuracy, but they do not independently test whether clients would actually contribute the predicted amounts, and the single-run results lack error bars or seed information.
- [Algorithm 1] Algorithm 1 initializes M over [1,...,L_max] even though the feasible range of L_c starts at L_min; the initialization should be for L_c = L_min,...,L_max.
Circularity Check
No significant circularity: the Nash and Stackelberg derivations follow algebraically from the paper's stated utility model; self-citations are not load-bearing.
full rationale
The central derivation is self-contained. Theorem 4's closed form (20) is obtained by solving the clients' first-order condition (16) and the aggregation identities (21)-(24); it is a mathematical consequence of the assumed utility functions, not an output that was used to define those utilities. Stage 2 substitutes this response into the owner's utility and optimizes, which is standard backward induction rather than a fitted-input prediction. The interior/boundary caveat attached to (19)-(20) ("where d*_n should satisfy the constraint of [0,Dn]") is a validity/domain concern about binding constraints, not a circular step, because the closed form is an algebraic solution of the stated game equations. The paper cites prior work by its own authors ([14], [15], [27]), but none of these citations smuggles in the target conclusion: [27] supplies the standard-function fixed-point theorem whose hypotheses (positivity, monotonicity, scalability) are verified in Theorem 2 for this game, so the uniqueness claim has an independent mathematical basis; [14]/[15] are used for privacy motivation and are supplemented by the paper's own SSIM experiments in Table III. The accuracy experiments allocate data according to the model's equilibrium formula, so they do not independently test the game model, but they are not used to derive the equilibrium and no fitted parameter is renamed as a prediction. Thus no step reduces to its own input by construction.
Assumptions & free parameters
free parameters (6)
- f_FLOPs regression coefficients (a,b) =
0.3779, -0.212
- |wc| regression coefficients (c,d) =
0.1098, 0.4711
- incentive weighting factors psi_n =
2.6e3 to 2.8e3 in experiments
- owner utility weights tau1, tau2 =
varied; e.g. ratio 4/6
- baseline utility offset S =
1e6 for Fig. 7c
- system and policy constants (d_req, Rmin, Rmax, Lmin, Lmax) =
2000, 60, 1000, 3, 12
assumptions (5)
- ad hoc to paper Client utility is quasi-linear: a proportional share of incentive minus linear energy cost, Eq. (12).
- domain assumption Owner utility is logarithmic in total contributed data, Eq. (2).
- ad hoc to paper f_FLOPs is affine and |wc| is exponential in Lc, Table II.
- standard math Standard-function theorem implies unique fixed point when its conditions hold.
- domain assumption Privacy leakage decreases with cut layer based on authors' prior work [14].
Cite this review
Pith. "Pith review of How Can Incentives and Cut Layer Selection Influence Data Contribution in Split Federated Learning?." pith.science (2026). https://pith.science/paper/XYRAXBA6
@misc{pith2026241207813,
author = {Pith},
title = {Pith review of: How Can Incentives and Cut Layer Selection Influence Data Contribution in Split Federated Learning?},
year = {2026},
howpublished = {\url{https://pith.science/paper/XYRAXBA6}},
note = {Machine review of arXiv:2412.07813}
}
read the original abstract
To alleviate the training burden in federated learning while enhancing convergence speed, Split Federated Learning (SFL) has emerged as a promising approach by combining the advantages of federated and split learning. However, recent studies have largely overlooked competitive situations. In this framework, the SFL model owner can choose the cut layer to balance the training load between the server and clients, ensuring the necessary level of privacy for the clients. Additionally, the SFL model owner sets incentives to encourage client participation in the SFL process. The optimization strategies employed by the SFL model owner influence clients' decisions regarding the amount of data they contribute, taking into account the shared incentives over clients and anticipated energy consumption during SFL. To address this framework, we model the problem using a hierarchical decision-making approach, formulated as a single-leader multi-follower Stackelberg game. We demonstrate the existence and uniqueness of the Nash equilibrium among clients and analyze the Stackelberg equilibrium by examining the leader's game. Furthermore, we discuss privacy concerns related to differential privacy and the criteria for selecting the minimum required cut layer. Our findings show that the Stackelberg equilibrium solution maximizes the utility for both the clients and the SFL model owner.
Figures
Figures from the paper (6 more)
Forward citations
Cited by 1 Pith paper
-
HASFL: Heterogeneity-aware Split Federated Learning over Edge Computing Systems
HASFL jointly optimizes per-device batch sizes and neural network split points to reduce training latency in heterogeneous split federated learning, guided by a new convergence bound.
Reference graph
Works this paper leans on
-
[1]
J. Lee, F. Solat, T. Y . Kim, and H. V . Poor, “Federated Learning- Empowered Mobile Network Management for 5G and Beyond Net- works: From Access to Core,” IEEE Communications Surveys & Tu- torials, vol. 26, no. 3, pp. 2176–2212, Jan. 2024. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 12
work page 2024
-
[2]
SplitFed: When Federated Learning Meets Split Learning,
C. Thapa, P. C. Mahawaga Arachchige, S. Camtepe, and L. Sun, “SplitFed: When Federated Learning Meets Split Learning,” in Proceed- ings of the 2022 Association for the Advancement of Artifical Intellgience (AAAI), 2022
work page 2022
-
[3]
Detailed Comparison of Communication Efficiency of Split Learning and Federated Learning,
A. Singh, P. Vepakomma, O. Gupta, and R. Raskar, “Detailed Comparison of Communication Efficiency of Split Learning and Federated Learning,” arXiv, 2019. [Online]. Available: https://arxiv.org/ abs/1909.09145
arXiv 2019
-
[4]
Split Federated Learning for 6G Enabled-Networks: Requirements, Challenges and Future Directions,
H. Hafi, B. Brik, P. A. Frangoudis, A. Ksentini, and M. Bagaa, “Split Federated Learning for 6G Enabled-Networks: Requirements, Challenges and Future Directions,” IEEE Access , vol. 12, pp. 9890– 9930, Jan. 2024
work page 2024
-
[5]
Accelerating Federated Learning with Split Learning on Locally Generated Losses,
D.-J. Han, H. I. Bhatti, J. Lee, and J. Moon, “Accelerating Federated Learning with Split Learning on Locally Generated Losses,” in Pro- ceedings of the 2021 International Conference on Machine Learning (ICML-Workshop of Federated Learning for User Privacy and Data Confidentiality), 2021
work page 2021
-
[6]
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 Journal on Selected Areas in Communications , vol. 41, no. 4, pp. 1051–1066, April 2023
work page 2023
-
[7]
Communication and Storage Efficient Federated Split Learning,
Y . Mu and C. Shen, “Communication and Storage Efficient Federated Split Learning,” arXiv, 2023. [Online]. Available: https: //arxiv.org/abs/2302.05599v1
arXiv 2023
-
[8]
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 Transactions on Mobile Computing , pp. 1–16, to appear, 2024
work page 2024
Show all 31 references
-
[9]
A Survey of In- centive Mechanism Design for Federated Learning,
Y . Zhan, J. Zhang, Z. Hong, L. Wu, P. Li, and S. Guo, “A Survey of In- centive Mechanism Design for Federated Learning,” IEEE Transactions on Emerging Topics in Computing, vol. 10, no. 2, pp. 1035–1044, 2022
2022
-
[10]
ESFL: Efficient Split Federated Learning over Resource-Constrained Heterogeneous Wireless Devices,
G. Zhu, Y . Deng, X. Chen, H. Zhang, Y . Fang, and T. F. Wong, “ESFL: Efficient Split Federated Learning over Resource-Constrained Heterogeneous Wireless Devices,” IEEE Internet of Things Journal , vol. 11, no. 16, pp. 27 153–27 166, 2024
2024
-
[11]
Split Federated Learning-Empowered Energy-Efficient Mobile Traffic Prediction Over UA Vs,
F. Solat, J. Lee, and D. Nyato, “Split Federated Learning-Empowered Energy-Efficient Mobile Traffic Prediction Over UA Vs,” IEEE Wireless Communications Letters, vol. 13, no. 11, pp. 3064–3068, 2024
2024
-
[12]
Multiple Classification with Split Learning,
J. Kim, S. Shin, Y . Yu, J. Lee, and K. Lee, “Multiple Classification with Split Learning,” in Proceedings of the 2020 International Conference on Smart Media and Applications (SMA) , 2020
2020
-
[13]
Reducing Leakage in Distributed Deep Learning for Sensitive Health Data,
P. Vepakomma, O. Gupta, A. Dubey, and R. Raskar, “Reducing Leakage in Distributed Deep Learning for Sensitive Health Data,” in Proceedings of the 2019 International Conference on Learning Representations (ICLR- AI for Social Good Workshop) , 2019
2019
-
[14]
Exploring the Privacy-Energy Consumption Tradeoff for Split Federated Learning,
J. Lee, M. Seif, J. Cho, and H. V . Poor, “Exploring the Privacy-Energy Consumption Tradeoff for Split Federated Learning,” IEEE Network , vol. 38, no. 6, pp. 388–395, May 2024
2024
-
[15]
Optimizing Privacy and Latency Tradeoffs in Split Federated Learning over Wireless Networks,
——, “Optimizing Privacy and Latency Tradeoffs in Split Federated Learning over Wireless Networks,” IEEE Wireless Communications Letters, pp. 1–1, to appear, 2024
2024
-
[16]
When MiniBatch SGD Meets SplitFed Learning:Convergence Analysis and Performance Evaluation,
C. Huang, G. Tian, and M. Tang, “When MiniBatch SGD Meets SplitFed Learning:Convergence Analysis and Performance Evaluation,” arXiv, 2023. [Online]. Available: https://arxiv.org/abs/2308.11953
2023 arXiv
-
[17]
User Preference Based Energy-Aware Mobile AR System with Edge Computing,
H. Wang and J. Xie, “User Preference Based Energy-Aware Mobile AR System with Edge Computing,” inProceedings of the 2020 IEEE Confer- ence on Computer Communications (INFOCOM), 2020, pp. 1379–1388
2020
-
[18]
An Edge Network Orches- trator for Mobile Augmented Reality,
Q. Liu, S. Huang, J. Opadere, and T. Han, “An Edge Network Orches- trator for Mobile Augmented Reality,” in Proceedings of the 2018 IEEE Conference on Computer Communications (INFOCOM), 2018, pp. 756– 764
2018
-
[19]
Market model and optimal pricing scheme of big data and Internet of Things,
D. Niyato, M. A. Alsheikh, P. Wang, D. I. Kim, and Z. Han, “Market model and optimal pricing scheme of big data and Internet of Things,” in in Proc. IEEE Int. Conf. Commun. (ICC) , May 2016, pp. 1–6
2016
-
[20]
HFEL: Joint Edge Association and Resource Allocation for Cost-Efficient Hierarchical Federated Edge Learning,
S. Luo, X. Chen, Q. Wu, Z. Zhou, and S. Yu, “HFEL: Joint Edge Association and Resource Allocation for Cost-Efficient Hierarchical Federated Edge Learning,” IEEE Transactions on Wireless Communi- cations, vol. 19, no. 10, pp. 6535–6548, 2020
2020
-
[21]
A Novel Joint Dataset and Computation Management Scheme for Energy-Efficient Federated Learning in Mobile Edge Computing,
J. Kim, D. Kim, J. Lee, and J. Hwang, “A Novel Joint Dataset and Computation Management Scheme for Energy-Efficient Federated Learning in Mobile Edge Computing,” IEEE Wireless Communications Letters, vol. 11, no. 5, pp. 898–902, 2022
2022
-
[22]
Joint Edge Server Selection and Dataset Management for Federated Learning-Enabled Mobile Traffic Prediction,
D. Doyeon, S. Shin, J. Jeong, and J. Lee, “Joint Edge Server Selection and Dataset Management for Federated Learning-Enabled Mobile Traffic Prediction,” IEEE Internet of Things Journal , vol. 11, no. 3, pp. 4971– 4986, Feb. 2024
2024
-
[23]
Federated Learning over Wireless Networks: Optimization Model De- sign and Analysis,
N. H. Tran, W. Bao, A. Zomaya, M. N. H. Nguyen, and C. S. Hong, “Federated Learning over Wireless Networks: Optimization Model De- sign and Analysis,” in Proceedings of the 2019 IEEE Conference on Computer Communications (INFOCOM) , 2019, pp. 1387–1395
2019
-
[24]
Basar and G
T. Basar and G. J. Olsder, Dynamic Noncooperative Game Theory, 2nd . Philadelphia, PA, USA: SIAM, 1999
1999
-
[25]
M. J. Osborne and A. Rubinstein, A Course in Game Theory . Cam- bridge, Jun. 1994
1994
-
[26]
Resource Pricing Game in Geo-distributed Clouds,
H. Roh, C. Jung, W. Lee, and D. Z. Du, “Resource Pricing Game in Geo-distributed Clouds,” in Proceedings of the 2013 IEEE Conference on Computer Communications (INFOCOM) , 2013, pp. 1519–1527
2013
-
[27]
Distributed Energy Trading in Microgrids: A Game-Theoretic Model and Its Equilibrium Analysis,
J. Lee, J. Guo, J. K. Choi, and M. Zukerman, “Distributed Energy Trading in Microgrids: A Game-Theoretic Model and Its Equilibrium Analysis,” IEEE Transactions on Industrial Electronics , vol. 62, no. 6, pp. 3524–3533, Jun. 2015
2015
-
[28]
Learning Multiple Layers of Features from Tiny Images,
A. Krizhevsky, G. Hinton et al., “Learning Multiple Layers of Features from Tiny Images,” Master’s thesis, University of Tronto, 2009
2009
-
[29]
Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms,
H. Xiao, K. Rasul, and R. V ollgraf, “Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms,” arXiv, 2017
2017
-
[30]
Decentralized Federated Learning Through Proxy Model Sharing,
S. Kalra, J. Wen, J. C. Cresswell, M. V olkovs, and H. R. Tizhoosh, “Decentralized Federated Learning Through Proxy Model Sharing,” Nature Communications, vol. 14, no. 2899, May 2023
2023
-
[31]
The Algorithmic Foundations of Differential Privacy,
C. Dwork, A. Roth et al., “The Algorithmic Foundations of Differential Privacy,” Foundations and Trends® in Theoretical Computer Science , vol. 9, no. 3–4, pp. 211–407, 2014
2014
Reviewed August 11, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.