acceptodds
Under review as a conference paper at ICLR 2027

PABP: An Approximately Truthful Auction Mechanism for Federated Learning with Bid Privacy Protection

Abstract

Auction mechanisms incentivize device participation in federated learning (FL), but existing designs typically require devices to disclose their private bids to the server. Protecting these bids is challenging because perturbed bids can distort allocation and payment decisions, thereby undermining economic incentives. We propose PABP, an approximately truthful auction mechanism that jointly designs bid perturbation, device selection, resource allocation, and payment determination. Each device locally perturbs its bid with Gaussian noise, and the server makes decisions using only the perturbed bids. PABP further approximates the theoretical Myerson payment under noisy bids, which provides -approximate incentive compatibility and individual rationality. Experiments across standard benchmarks and diverse data settings show that PABP protects bid privacy while consistently outperforming representative baselines.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Auction, bid protection, Approximation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_DjLaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PABP selects FL participants using locally perturbed bids and calibrates payments toward an expected truthful-payment target. The allocation analysis is sound under the stated model, and experiments retain high welfare. The practical incentive and privacy guarantees need quantitative support.

Critical strengths
  1. The formulation accounts for device resource costs and separates optimal allocation for perturbed bids from welfare evaluated at true values.
  2. The appendix gives a formal single-release privacy analysis and clearly distinguishes finite-sample payment audits from the worst-case error required by the theorem.
Critical weaknesses
  1. Approximate truthfulness and participation depend on a uniform payment-error bound that the proposed calibration does not establish. Twelve calibration profiles and twenty validation profiles cannot certify all bids and opponent profiles. Moreover, noise can select a zero-value device that incurs resource costs, while executable service payments are nonnegative. Some participation loss is therefore unavoidable; the paper needs a meaningful bound on its magnitude.
  2. The experiments do not connect the chosen noise scale to a specified privacy target. The flat reconstruction-error curve uses repeated copies of one noisy bid, as the appendix explains. This supports the fixed-bid experiment but leaves protection under changing bids unresolved. Accuracy comparisons with methods protecting model updates also do not isolate the bid-privacy trade-off.
Questions for authors
  1. Can the authors bound worst-case payment error for the executable rule, including calibration uncertainty, and express resulting deviation gains and participation losses relative to device values?
  2. What privacy radius and target does the experimental noise scale achieve? How do welfare and incentive errors change at stronger privacy levels and with multiple distinct releases?
3 comments
AuthorsOpenReview ↗

Weakness and Concerns (1/3)

We thank the reviewer for the helpful comments.

Weakness 1 & Question 1:

Q: Can the authors bound worst-case payment error for the executable rule, including calibration uncertainty, and express resulting deviation gains and participation losses relative to device values? R: Thank you for this question. We fully agree that a payment-error upper bound should be provided, and the corresponding maximum expected deviation gain and expected participation loss should be analyzed. In the following, we have expanded the theoretical analysis to address the expected deviation gain and loss.

Uniform payment-error bound. Under the revised payment-calibration constraints described in Section 4.3 and details in Appendix A.5, the payment-error analysis in Appendix B.4 establishes the following bound, also stated in Section 4.4.

δi=sup⁡b∈BN∣Eη[πi(b~;θi∗)∣b]−πiref(bi;b−i)∣≤αiσi2π.\delta_i=\sup_{b\in B^N}\left\lvert E_{\eta}\left[\pi_i(\tilde{b};\theta_i^{*})\mid b\right]-\pi_i^{ref}(b_i;b_{-i})\right\rvert\le\frac{\alpha_i\sigma_i}{\sqrt{2\pi}}.

The bound holds for every adopted calibration output satisfying the continuous-domain constraints. Thus, although finite random samples are used for fitting, the guarantee does not depend on their coverage of the bid space.

Deviation gain and participation loss.

With the approximate expected IC/IR guarantees in Section 4.4 (Theorem 1), the maximum expected utility gain from misreporting is at most εIC,i=2δi=2π αiσi\varepsilon_{IC,i}=2\delta_i=\sqrt{\frac{2}{\pi}}\,\alpha_i\sigma_i, and the expected participation loss is at most −εIR,i=−δi=−αiσi2π-\varepsilon_{IR,i}=-\delta_i=-\frac{\alpha_i\sigma_i}{\sqrt{2\pi}}.

We also respond to the concern in Weakness 1 regarding participation loss when noise selects a zero-valuation device. We clarify the example of a zero-valuation device winning after bid perturbation. This outcome is rare because the noise is unlikely to raise a zero bid to the selection threshold. In practice, the bid required for selection is much larger than the noise scale. Furthermore, from the individual rationality standpoint, a device, though selected due to noise, will not choose to train the model or pay for a zero-valued service. For example, a customer would not purchase an unwanted product merely because a shopkeeper offered it. Similarly, selected devices with zero valuation are unwilling to pay in this case.

AuthorsOpenReview ↗

Weakness and Concerns (2/3)

Question 2:

Q: What privacy radius and target does the experimental noise scale achieve? How do welfare and incentive errors change at stronger privacy levels and with multiple distinct releases?

R: Thank you for requesting a quantitative connection between the experimental noise scale, the privacy target, and the resulting welfare and incentive errors. We have quantified the cumulative privacy budget with the experimental noise scale, and added two experiments to measure how stronger privacy and more biddings affect social welfare, payment, IC, and IR errors.

Privacy radius and target. In our paper, we use ρ-zero-concentrated differential privacy (ρ-zCDP) for privacy protection. Let B=[0,Bbid]B=[0,{B}^{bid}] be the bid interval, where Bbid{B}^{bid} is the bid upper bound. For device ii, consider two bid sequences for TT training rounds, i.e., b=(bi(1),…,bi(T))b=(b_i^{(1)},\ldots,b_i^{(T)}) and b′=(bi′(1),…,bi′(T))b'=(b_i^{\prime(1)},\ldots,b_i^{\prime(T)}). Here, t∈{1,…,T}t \in \lbrace 1,\ldots,T\rbrace indexes these ordinary rounds, but refresh rounds are excluded. In our experiments, the privacy radius (per-release sensitivity) is Δb=Bbid=1\Delta_b=B^{bid}=1. The sequences are neighboring if ∣bi(t)−bi′(t)∣≤Δb\lvert b_i^{(t)}-b_i^{\prime(t)}\rvert\le\Delta_b for every tt.

The privacy target (also called privacy budget) is the cumulative ρ-zCDP budget under the values of the noise scale and privacy radius. With independent Gaussian noise of scale σ at each round, the cumulative privacy budget is defined as

ρ=TΔb22σ2.\rho =\frac{T{\Delta }_{b}^{2}}{2{\sigma }^{2}}.

Based on the above descriptions, Table 5 illustrates the noise scale and zCDP budget under different numbers of bid submission rounds.

Table 5. Cumulative zCDP privacy budgets for 1 and 80 ordinary-round bid submissions.

Bid submission in ordinary-roundsensitivityNoise standard deviationCumulative zCDP budget
110.1234.722
8010.122777.778

Welfare and incentive errors. We added two controlled auction experiments with N=20N=20, K=5K=5, and five independent repetitions to evaluate stronger privacy and more perturbed-bid releases.

We evaluate the following ratio of expected true social welfare achieved by PABP to the optimum without bid perturbation. This ratio measures how much of the optimal social welfare PABP retains while protecting bids. A higher ratio indicates better welfare preservation, as defined in Appendix B.5.

ΓSW=∑ℓ∉RE[SW(ℓ)]∑ℓ∉RSWFI∗,(ℓ).{\Gamma }_{SW}=\frac{\sum _{\ell \notin R} E[S{W}^{(\ell )}]}{\sum _{\ell \notin R} S{W}_{FI}^{*,(\ell )}}.

PABP and the optimum are evaluated under the same devices, true valuations, and resource conditions in each round.

Stronger privacy levels. To evaluate how stronger privacy affects welfare and incentive errors, we fix the number of bid submissions at 80 and increase the noise scale σ\sigma from 0.06 to 0.30. This reduces the cumulative privacy budget and strengthens privacy.

Multiple distinct perturbed-bid releases. We fix σ=0.12\sigma=0.12 and the single-release privacy budget at Δb2/(2σ2)=34.722222\Delta_b^2/(2\sigma^2)=34.722222. The total release count up to the current round is T∈{1,10,40,80}T\in\{1,10,40,80\}. All groups use the same bank of 80 current-round bid scenarios, quality coefficients, CPU frequencies, upload powers, resource costs, calibration parameters, and evaluation random numbers. Only the number of prior releases changes. Each prior release uses a new bid and fresh independent noise.

Table 6. Effects of stronger privacy and more bid releases on social welfare and payment, IC, and IR errors.

ConditionsSocial welfare ratio (ΓSW)({\Gamma }_{SW})Payment error (δival)({\delta }_{i}^{val})IC error (εIC,i)({\varepsilon }_{IC,i})IR error (εIR,i)({\varepsilon }_{IR,i})
Stronger privacy: σ = 0.06 → 0.30 (T=80T= 80 fixed)99.39403% → 88.39403%0.99644% → 6.17883%0.09046% → 1.24104%0.02440% → 0.81467%
More bid releases: T=1→80T=1 \rightarrow 80 (ρ = 34.72222234.722222, σ=0.12\sigma=0.12fixed)97.62691% → 97.62753%2.23381% → 2.23243%0.32367% → 0.32367%0.11693% → 0.11660%
AuthorsOpenReview ↗

Weakness and Concerns (3/3)

Results in Table 6 confirm lower welfare retention and higher economic errors under stronger privacy. This is reasonable, because stronger privacy means more noise to the submitted bids, which can severely break the economic properties. Moreover, under multiple training rounds of bid submission, the social welfare ratio, payment error, IC error and IR error remain nearly unchanged. The minor fluctuations are caused by random noise on perturbed bids and are expected. This demonstrates that our method works properly under multiple distinct releases.

Weakness 2:

W: The experiments do not connect the chosen noise scale to a specified privacy target. The flat reconstruction-error curve uses repeated copies of one noisy bid, as the appendix explains. This supports the fixed-bid experiment but leaves protection under changing bids unresolved. Accuracy comparisons with methods protecting model updates also do not isolate the bid-privacy trade-off.

R: We have added an experiment to evaluate whether bid privacy protection is still valid under multiple biddings.

Dynamic-bid privacy protection experiment. Each device draws a new valuation independently from a uniform distribution on [0,1] in every ordinary round, and perturbs it with fresh Gaussian noise of standard deviation 0.12. The server knows the bid distribution and perturbation mechanism, and uses the available history to reconstruct the device’s true bid.

Table 1. Current true-bid reconstruction error (RMSE) after 80 ordinary rounds: PABP versus the auction without bid privacy (Appendix C.1).

ConfigurationPABP (RMSE)Auction without bid privacy (RMSE)
MNIST0.105268 ± 0.0030400
Fashion-MNIST0.105654 ± 0.0028310
CIFAR-100.104622 ± 0.0035690

The root mean square error (RMSE) measures the discrepancy between reconstructed and original bids. A value of zero means exact bid recovery. A positive value indicates imperfect recovery and some privacy protection against the evaluated attack. PABP’s mean RMSE remains 0.1041–0.1094 across 1–80 ordinary-round releases, compared with zero for the Non-private auction, demonstrating that the proposed PABP can protect the recovery of current true bids under dynamic bidding.

Baseline settings.

We note that PrivAim [1] and RAFL [2] also perturb participants’ bids bb for privacy protection. Therefore, we add experiments on the ImageNet dataset to compare PABP with PrivAim-bid-adapted and RAFL-bid-adapted.

Our comparison metric is the social welfare ratio, defined as the ratio of the aggregated expected true social welfare to the full-information optimal social welfare without bid perturbation. This ratio measures how much of the optimal social welfare each method retains while protecting bids. A higher ratio indicates better welfare preservation and less social welfare loss.

Table 2. Social welfare ratios of PABP, PrivAim-bid-adapted, and RAFL-bid-adapted.

MethodSocial welfare ratio
PABP97.215%
PrivAim-bid-adapted88.129%
RAFL-bid-adapted87.560%

Table 2 shows that PABP achieves a higher social welfare ratio and less social welfare loss than both adapted baselines.

Based on our analysis, we believe that PABP’s additional resource optimization contributes to this improvement. PABP jointly considers winner selection and resource allocation. It computes optimal CPU frequencies and upload powers and assigns these configurations to the selected devices, reducing their resource costs. PrivAim and RAFL do not optimize CPU frequencies or upload powers.

[1] Wang et al. (2023). PrivAim: A Dual-Privacy Preserving and Quality-Aware Incentive Mechanism for Federated Learning. IEEE Transactions on Computers, 72(7):1913–1927. DOI.

[2] Sun et al. (2026). RAFL: A Reverse Auction Federated Learning Framework with Non-Independent and Identically Distributed Data for Mobile Crowdsensing. Computer Networks, 283:112314. DOI.

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

This paper proposes PABP, a federated learning auction that uses locally perturbed bids to select devices, allocate resources, and determine service payments. It calibrates payments toward an expected Myerson target and derives conditional approximate incentive compatibility and individual rationality guarantees.

Critical strengths
  1. The paper addresses a relevant problem by jointly considering bid privacy, participant selection, resource allocation, and payments in resource-constrained federated learning.
  2. The formulation distinguishes true welfare from the perturbed-bid allocation objective and explicitly incorporates device training costs, communication costs, and deadline constraints.
Critical weaknesses
  1. Payment calibration lacks uniform error guarantees. The theorem assumes bounded worst-case conditional expected payment error, but calibration covers only finitely sampled bid profiles, leaving the implemented IC/IR bounds unestablished.
  2. Nonnegative payments create irreducible approximation error. Zero-valuation devices may win under noise and incur costs, requiring server compensation for expected IR. Yet payments are nonnegative and calibration only raises charges, so more samples cannot eliminate this error.
  3. The baselines address different privacy and incentive settings: IAFL uses no monetary bids, while QI-DPFL protects model updates. Accuracy gains therefore do not isolate bid-protection benefits; equivalent auction and privacy controls are missing.
  4. Repeatedly observing the same perturbed bid does not test dynamic-bid privacy. Stable reconstruction errors provide no evidence about cumulative leakage when bids change or attacks exploiting allocation and payment information.
  5. Finite-bid payment checks and limited utility curves cannot establish worst-case IC/IR bounds across the full bid space.
Questions for authors

See Weakness

4 comments
AuthorsOpenReview ↗

Weakness and Concerns (1/3)

We thank the reviewer for the following five concerns.

Weakness 1:

W: Payment calibration lacks uniform error guarantees. The theorem assumes bounded worst-case conditional expected payment error, but calibration covers only finitely sampled bid profiles, leaving the implemented IC/IR bounds unestablished.

R: Thank you for mentioning the lack of error guarantees of devised payment. We agree that an error bound of payment rule can contribute to establishing the approximate IC/IR guarantees. Following your suggestion, we have added a payment-error bound in Section 4.4, with its proof in Appendix B.4. Specifically, with the payment-calibration constraints described in Section 4.3 and Appendix A.5, we prove that

δi=sup⁡b∈BN∣Eη[πi(b~;θi∗)∣b]−πiref(bi;b−i)∣≤αiσi2π.{\delta }_{i}=\sup_{b\in {B}^{N}} \left\lvert {E}_{\eta }\left[ {\pi }_{i}(\tilde{b};{\theta }_{i}^{*})\mid b\right]-{\pi }_{i}^{ref}({b}_{i};{b}_{-i})\right\rvert\le \frac{{\alpha }_{i}{\sigma }_{i}}{\sqrt{2\pi }}.

Based on the payment error bound, i.e., δi\delta_i, we have the approximate IC and IR guarantees respectively as

εIC,i=2δi=2π αiσi.\varepsilon_{IC,i}=2\delta_i=\sqrt{\frac{2}{\pi}}\,\alpha_i\sigma_i.

−εIR,i=−δi=−αiσi2π.-\varepsilon_{IR,i}=-\delta_i=-\frac{\alpha_i\sigma_i}{\sqrt{2\pi}}.

Moreover, in the paper, we state that calibration profiles are sampled from a finite bid space BN,{B}^{N}, where B=[0,Bbid]B=[0,{B}^{bid}] is defined in Section 3.1. This is because from the perspective of auction theory, each single client's bid should be within a scope [1] [2] [3].

[1] Li et al. (2024). Performance-Based Pricing for Federated Learning via Auction. PVLDB, 17(6):1269–1282.

Weakness 2:

W: Nonnegative payments create irreducible approximation error. Zero-valuation devices may win under noise and incur costs, requiring server compensation for expected IR. Yet payments are nonnegative and calibration only raises charges, so more samples cannot eliminate this error.

R: Thank you for pointing out that the original nonnegative correction could only increase charges and that additional samples could not eliminate the resulting participation loss. We fully agree and have revised the payment calibration accordingly.

We next clarify the example of a zero-valuation device winning after bid perturbation. This outcome is rare because the noise is unlikely to raise a zero bid to the level needed for selection. Holding other conditions fixed, the bid required for selection is much larger than the noise scale. Furthermore, from the individual rationality standpoint, a device will not choose to participate or pay for a zero-valued service even when picked by noise. For example, a customer would not purchase an unwanted product merely because a shopkeeper offered it. Similarly, selection does not create a willingness to pay in a zero-valuation device.

We apologize for the confusion caused by our description of the payment correction. The statement that calibration could only increase charges did not accurately describe the payment implementation used in our experiments. The adjustment relative to the original payment can be positive or negative: calibration can lower charges, while the reduction cannot exceed the original payment. The final payment remains nonnegative.

We have revised the description to distinguish the payment correction from the nonnegative final payment. The added controlled experiment also measures whether calibration reduces payment approximation error.

Table 4. Mean validation payment errors before and after calibration at two noise scales.

Noise scaleUncalibrated payment errorCalibrated payment errorRelative reduction
0.122.904%2.231%23.16%
0.307.423%6.179%16.76%

This experiment compares payment errors before and after calibration. Calibration reduces the average error under both noise scales.

AuthorsOpenReview ↗

Weakness and Concerns (2/3)

Weakness 3:

W: The baselines address different privacy and incentive settings: IAFL uses no monetary bids, while QI-DPFL protects model updates. Accuracy gains therefore do not isolate bid-protection benefits; equivalent auction and privacy controls are missing.

R: Thank you for highlighting the differences between the baseline settings. We note that PrivAim [1] and RAFL [2] also perturb participants’ bids bb for privacy protection. Therefore, we add experiments on the ImageNet dataset to compare PABP with PrivAim-bid-adapted and RAFL-bid-adapted. Our comparison metric is the social welfare ratio, defined as the ratio of the aggregated expected true social welfare to the full-information optimal social welfare without bid perturbation. This ratio measures how much of the optimal social welfare each method retains while protecting bids. A higher ratio indicates better welfare preservation and less social welfare loss.

Table 2. Social welfare ratios of PABP, PrivAim-bid-adapted, and RAFL-bid-adapted.

MethodSocial welfare ratio
PABP97.215%
PrivAim-bid-adapted88.129%
RAFL-bid-adapted87.560%

Table 2 shows that PABP achieves a higher social welfare ratio and less social welfare loss than both adapted baselines. Based on our analysis, we believe that PABP’s additional resource optimization contributes to this improvement. PABP jointly considers winner selection and resource allocation. It computes optimal CPU frequencies and upload powers and assigns these configurations to the selected devices, reducing their resource costs. PrivAim and RAFL do not optimize CPU frequencies or upload powers.

[1] Wang et al. (2023). PrivAim: A Dual-Privacy Preserving and Quality-Aware Incentive Mechanism for Federated Learning. IEEE Transactions on Computers, 72(7):1913–1927. DOI.

[2] Sun et al. (2026). RAFL: A Reverse Auction Federated Learning Framework with Non-Independent and Identically Distributed Data for Mobile Crowdsensing. Computer Networks, 283:112314. DOI.

Weakness 4:

W: Repeatedly observing the same perturbed bid does not test dynamic-bid privacy. Stable reconstruction errors provide no evidence about cumulative leakage when bids change or attacks exploiting allocation and payment information.

R: Thank the reviewers for pointing out this issue. We agree that the original experiment considered a fixed true bid and did not adequately evaluate whether bid privacy protection is still valid under multiple biddings, where a device submits a new bid with fresh noise in each ordinary round. To this end, in the revised manuscript, we have established an explicit upper bound on the cumulative privacy budget under multiple biddings (Appendix A.6.1).

ρi(T)≤∑ℓ∈OLΔb22(σi(ℓ))2=TΔb22σi2whenσi(ℓ)=σi.{\rho }_{i}^{(T)}\le \sum _{\ell \in {O}_{L}} \frac{{\Delta }_{b}^{2}}{2\bigl({\sigma }_{i}^{(\ell )}\bigr)^{2}}=\frac{T{\Delta }_{b}^{2}}{2{\sigma }_{i}^{2}}\quad \text{when}\quad { \sigma }_{i}^{(\ell )}={\sigma }_{i}.

Here, Δb\Delta_b is the bid sensitivity, which is the maximum permitted difference between neighboring true bids in each ordinary round. σi\sigma_i is the noise scale for device ii, namely the standard deviation of the Gaussian noise. T=∣OL∣T=\lvert O_L\rvert is the number of ordinary rounds in which device ii submits a new perturbed bid. The privacy budget ρi(T)\rho_i^{(T)} characterizes the extent to which the server can distinguish between two neighboring true-bid sequences after observing allocations and payments, thereby reflecting its ability to infer true bids. A smaller privacy budget provides a stronger privacy guarantee. We establish an upper bound on the cumulative privacy budget across multiple rounds. This bound shows that the server’s ability to infer true bids is explicitly constrained.

Furthermore, allocations and payments are computed from protected, perturbed bids. Devices' allocation and payment information therefore also falls within the scope of privacy protection, and there is no cumulative leakage when bids change or when the server launches attacks exploiting allocation and payment information. We have added an experiment to evaluate whether bid privacy protection is still valid under multiple biddings.

AuthorsOpenReview ↗

Weakness and Concerns (3/3)

Dynamic-bid privacy protection experiment. Each device draws a new valuation independently from a uniform distribution on [0,1] in every ordinary round, and perturbs it with fresh Gaussian noise of standard deviation 0.12. The server knows the bid distribution and perturbation mechanism, and uses the available history to reconstruct the device’s true bid.

Table 1. Current true-bid reconstruction error (RMSE) after 80 ordinary rounds: PABP versus the auction without bid privacy (Appendix C.1).

ConfigurationPABP (RMSE)Auction without bid privacy (RMSE)
MNIST0.105268 ± 0.0030400
Fashion-MNIST0.105654 ± 0.0028310
CIFAR-100.104622 ± 0.0035690

The root mean square error (RMSE) measures the discrepancy between reconstructed and original bids. A value of zero means exact bid recovery. A positive value indicates imperfect recovery and some privacy protection against the evaluated attack. PABP’s mean RMSE remains 0.1041–0.1094 across 1–80 ordinary-round releases, compared with zero for the Non-private auction, demonstrating that the proposed PABP can protect the recovery of current true bids under dynamic bidding.

Weakness 5:

W: Finite-bid payment checks and limited utility curves cannot establish worst-case IC/IR bounds across the full bid space.

R: Thank you for highlighting Finite-bid payment checks and limited utility curves. We have established a uniform payment-error bound and the resulting approximate expected IC and IR guarantees over the full admissible bid space in Section 4.4, with proofs in Appendix B.4. The finite-sample payment checks and utility curves are used to validate these theoretical conclusions experimentally.

Theoretical guarantees. The payment-error analysis in Appendix B.4 establishes

δi=sup⁡b∈BN∣Eη[πi(b~;θi∗)∣b]−πiref(bi;b−i)∣≤αiσi2π.{\delta }_{i}=\sup_{b\in {B}^{N}} \left\lvert {E}_{\eta }[{\pi }_{i}(\tilde{b};{\theta }_{i}^{*})\mid b]-{\pi }_{i}^{ref}({b}_{i};{b}_{-i})\right\rvert\le \frac{{\alpha }_{i}{\sigma }_{i}}{\sqrt{2\pi }}.

The approximate expected IC/IR guarantees in Section 4.4 (Theorem 1), derived in Appendix B.4.5, therefore establish

UiE(vi,bi′;b−i)−UiE(vi,vi;b−i)≤εIC,i=2δi=2π αiσi,{U}_{i}^{E}({v}_{i},{b}_{i}';{b}_{-i})-{U}_{i}^{E}({v}_{i},{v}_{i};{b}_{-i})\le {\varepsilon }_{IC,i}={2\delta }_{i}=\sqrt{\frac{2}{\pi }}\,{\alpha }_{i}{\sigma }_{i},

and

UiE(vi,vi;b−i)≥−εIR,i=−δi=−αiσi2π.{U}_{i}^{E}({v}_{i},{v}_{i};{b}_{-i})\ge -{\varepsilon }_{IR,i}=-{\delta }_{i}=-\frac{{\alpha }_{i}{\sigma }_{i}}{\sqrt{2\pi }}.

These guarantees hold for all admissible valuations, reports, and opponent bid profiles. We prove theoretical upper bounds for the payment calibration error and approximate IC/IR error over the full bid space. Finite samples are only used for payment error calibration computation and simulation experiments to verify this theoretical result.

Dear Reviewer CStL,

We are approaching the end of the discussion phase, and we have unfortunately received no feedback from you on our rebuttal.

Can you please kindly take a look at our responses, and let us know whether we have clarified your concerns? Thank you so much if you are available.

Paper28078 Authors

Reviewer_DXRnOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PABP, a privacy-aware auction mechanism tailored for federated learning. PABP aims to protect participants' private bids while maintaining allocation efficiency and economic incentives. Specifically, PABP employs local Gaussian perturbation to protect original bids and jointly optimizes device selection and computing and communication resource allocation based on the perturbed bids. To address the potential impact of bid perturbation on incentive compatibility, the authors introduce a Myerson-Calibrated Payment mechanism, which approximates the ideal expected payment using only perturbed bid information. The work includes theoretical analyses of allocation optimality, expected allocation monotonicity, and approximate incentive compatibility and individual rationality under bounded payment approximation errors. Evaluations on MNIST, Fashion-MNIST, and CIFAR-10 show that PABP delivers competitive learning performance relative to its non-private auction baseline and retains roughly 97%–98% of the full-information optimal social welfare.

Critical strengths

1、This paper addresses a meaningful problem by jointly considering bid privacy protection, device selection, resource allocation, and payment determination in federated learning. PABP enables the server to make allocation and payment decisions using only perturbed bids, without directly accessing the participants' original bids. 2、The paper provides a relatively comprehensive theoretical analysis, including allocation optimality, expected allocation monotonicity, and conditional approximate incentive compatibility and individual rationality guarantees. In particular, the proposed Myerson-Calibrated Payment mechanism attempts to bridge the gap between the ideal expected payment and an executable payment rule based solely on perturbed bids. 3、Experimental results show that PABP achieves learning accuracy comparable to the non-private auction baseline across MNIST, Fashion-MNIST, and CIFAR-10. Furthermore, the reported social welfare retention of approximately 97%–98% suggests that the proposed bid perturbation mechanism can preserve much of the allocation efficiency under the evaluated settings. The additional ablation studies and incentive compatibility evaluations also help clarify the behavior of the proposed PABP .

Critical weaknesses

1、The algorithmic novelty of the joint allocation method is limited, and the distinction from existing mechanism-design approaches is insufficiently justified. The proposed joint device selection and resource allocation problem can be decomposed into independent minimum-cost resource optimization for each device, followed by a cardinality-constrained selection problem. Specifically, after computing each device's minimum resource cost, the selection objective in Eq. (16) reduces to maximizing a sum of independent device scores under a cardinality constraint. This problem admits a straightforward optimal solution by selecting the highest positive scores, up to the capacity limit K. Therefore, the proposed branch-and-bound algorithm appears unnecessarily complicated for the current problem formulation. Appendix C.8 further confirms that the branch-and-bound solutions exactly match those obtained by direct Top-K selection.

2、The empirical results do not sufficiently substantiate the claimed advantages of the proposed auction mechanism. The main comparisons are conducted against Non-private auction, IAFL-adapted, and QI-DPFL-adapted. However, these baselines differ substantially in their underlying objectives and privacy protection strategies. In particular, IAFL-adapted does not incorporate monetary bids, while QI-DPFL-adapted primarily perturbs model updates rather than protecting bid information. In contrast, PABP perturbs only the bids without directly modifying the model updates. Therefore, its superior test accuracy over these baselines does not necessarily demonstrate the superiority of the proposed privacy-preserving auction mechanism. More directly comparable baselines are needed to isolate and validate the benefits of the proposed approach. Additionally, the experiments are limited to relatively small-scale benchmark datasets, including MNIST, Fashion-MNIST, and CIFAR-10, using VGG9 as the global model. The paper leaves unaddressed how PABP fares on larger datasets such as ImageNet or with deeper, more complex network architectures. Additional experiments under more realistic settings would strengthen the empirical evidence for the proposed PABP .

3、The related work section does not adequately discuss several existing studies closely related to the core contributions of PABP, particularly those investigating privacy protection and incentive mechanisms in federated learning. Relevant studies include:  PrivAim: A Dual-Privacy Preserving and Quality-Aware Incentive Mechanism for Federated Learning.  BIT-FL: Blockchain-Enabled Incentivized and Secure Federated Learning Framework.  RAFL: A Reverse Auction Federated Learning Framework with Non-Independent and Identically Distributed Data for Mobile Crowdsensing. The lack of a systematic comparison with these studies makes it difficult to assess the novelty of PABP and to identify the specific limitations of existing approaches that the proposed framework addresses. In particular, the authors should clarify how the proposed Gaussian bid perturbation, Myerson-Calibrated Payment, and joint resource allocation differ from, or improve upon, existing privacy-preserving incentive mechanisms. A more comprehensive discussion of closely related work, highlighting both the technical distinctions and the specific advances of PABP, is necessary to establish the paper's novelty and significance.

Questions for authors
  1. Given that the device selection problem in Eq. (16) can be optimally solved by directly selecting the Top-K devices with positive scores, what is the motivation for adopting a branch-and-bound algorithm? The authors are asked to clarify the specific technical advantages of the proposed approach over a straightforward Top-K selection strategy, particularly in terms of computational complexity, efficiency, and scalability?
  2. Could the authors provide comparisons with more directly relevant privacy-preserving auction mechanisms that protect bid information under comparable settings?
  3. Could the authors further evaluate the scalability of PABP with larger datasets (e.g., ImageNet), deeper neural network architectures, and scenarios with growing number of candidate devices?
  4. Could the authors clarify the key technical differences between PABP and closely related studies, such as PrivAim, BIT-FL, and RAFL? Specifically, what are the specific novel contributions of the proposed Gaussian bid perturbation, Myerson-Calibrated Payment, and joint resource allocation compared with existing privacy-preserving incentive mechanisms?
7 comments
AuthorsOpenReview ↗

Weakness and Concerns (1/3)

We thank the reviewer for the constructive feedbacks.

Weakness 1 & Question 1:

Q: Given that the device selection problem in Eq. (16) can be optimally solved by directly selecting the Top-K devices with positive scores, what is the motivation for adopting a branch-and-bound algorithm? The authors are asked to clarify the specific technical advantages of the proposed approach over a straightforward Top-K selection strategy, particularly in terms of computational complexity, efficiency, and scalability?

R: Thank you for asking us to clarify the motivation for branch-and-bound and its computational complexity, efficiency, and scalability relative to Top-K selection.

Using protected bids, we view branch-and-bound as a new approach to optimizing allocation variables such as the device-selection variable xx within this joint-allocation framework. Binary participation means that each device is either selected or not selected, which naturally corresponds to inclusion and exclusion branches. The objective is the sum of the selected devices’ net scores. This lets us bound the best result each branch can achieve. We stop searching a branch if it cannot improve the current best solution.

Scalability. As the number of candidate devices increases, each additional device introduces a resource subproblem of the same form. These involve optimizing the device’s CPU frequency, upload power, local computation resource allocation, and local storage quota. During the search for device combinations, the branch-and-bound algorithm calculates resource usage and checks whether resource limits are exceeded. It prunes unpromising branches in advance instead of recomputing the resource allocation for all devices for every candidate combination.

Our branch-and-bound scheme can also accommodate future settings with additional constraints, especially combinatorial constraints among devices. For example, total bandwidth or total upload-slot constraints can be introduced when the system has shared resource limits. A device’s eligibility for selection then depends on how much shared resources other devices consume. Therefore, ranking devices by their individual net scores using Top-K is generally no longer applicable.

Complexity and efficiency. Direct Top-K sorting takes O(n log n), once resource-adjusted net scores are available. This excludes the continuous optimization required by both methods to obtain each device’s minimum resource cost, CPU frequency, and upload power. The comparison should therefore consider the complete joint-allocation procedure and its ability to handle constraints, rather than sorting time alone. Under the current constraints, each device’s resource problem can be solved independently. We compute its minimum resource cost and the corresponding CPU frequency and upload power once. These results are stored before the branch-and-bound search. The search then involves only binary participation decisions, and different branches reuse the stored costs and configurations. Objective bounds prune combinations that cannot improve the current best solution, avoiding repeated continuous optimization at search nodes. Once the device set is selected, its stored CPU frequencies and upload powers are assigned. This structure reduces repeated computation and supports the constraint extensions above without implying a lower worst-case complexity than sorting.

Weakness 2 & Question 2:

Q: Could the authors provide comparisons with more directly relevant privacy-preserving auction mechanisms that protect bid information under comparable settings?

R: Thank you for requesting comparisons with bid-private auction mechanisms under comparable settings.

We note that PrivAim [1] and RAFL [2] also perturb participants’ bids bb for privacy protection. Therefore, we add experiments on the ImageNet dataset to compare PABP with PrivAim-bid-adapted and RAFL-bid-adapted. Our comparison metric is the social welfare ratio, defined as the ratio of the aggregated expected true social welfare to the full-information optimal social welfare without bid perturbation. This ratio measures how much of the optimal social welfare each method retains while protecting bids. A higher ratio indicates better welfare preservation and less social welfare loss.

Table 2. Social welfare ratios of PABP, PrivAim-bid-adapted, and RAFL-bid-adapted.

MethodSocial welfare ratio
PABP97.215%
PrivAim-bid-adapted88.129%
RAFL-bid-adapted87.560%
AuthorsOpenReview ↗

Weakness and Concerns (2/3)

Table 2 shows that PABP achieves a higher social welfare ratio and less social welfare loss than both adapted baselines. Based on our analysis, we believe that PABP’s additional resource optimization contributes to this improvement. PABP jointly considers winner selection and resource allocation. It computes optimal CPU frequencies and upload powers and assigns these configurations to the selected devices, reducing their resource costs. PrivAim and RAFL do not optimize CPU frequencies or uploading powers.

[1] Wang et al. (2023). PrivAim: A Dual-Privacy Preserving and Quality-Aware Incentive Mechanism for Federated Learning. IEEE Transactions on Computers, 72(7):1913–1927. DOI.

[2] Sun et al. (2026). RAFL: A Reverse Auction Federated Learning Framework with Non-Independent and Identically Distributed Data for Mobile Crowdsensing. Computer Networks, 283:112314. DOI.

Weakness 2 & Question 3:

Q: Could the authors further evaluate the scalability of PABP with larger datasets (e.g., ImageNet), deeper neural network architectures, and scenarios with growing number of candidate devices?

R: Thank you for suggesting evaluation on larger datasets. We have conducted an additional four-method comparison on ImageNet-1K with 20 devices and 60 global training rounds. All methods use the same ImageNet-1K split and common training settings. The results are summarized below.

Table 7. Top-1 and Top-5 test accuracy and server-visible bid information on ImageNet-1K.

MethodBid visible to serverTop-1 test accuracy (%)Top-5 test accuracy (%)
PABP (Ours)Noise-perturbed bid51.35477.242
Non-private auctionTrue bid51.84877.632
IAFL-adaptedNo monetary bid22.04645.630
QI-DPFL-adaptedTrue bid and perturbed model update2.2611.30

On ImageNet-1K, PABP’s Top-1 and Top-5 test accuracies are 0.494 and 0.390 percentage points below the Non-private auction, respectively, and exceed those of IAFL-adapted and QI-DPFL-adapted. These results show that PABP also maintains test accuracy close to the Non-private auction on ImageNet-1K. Meanwhile, the original valuation bids remain protected.

AuthorsOpenReview ↗

Weakness and Concerns (3/3)

Weakness 3 & Question 4:

Q: Could the authors clarify the key technical differences between PABP and closely related studies, such as PrivAim, BIT-FL, and RAFL? Specifically, what are the specific novel contributions of the proposed Gaussian bid perturbation, Myerson-Calibrated Payment, and joint resource allocation compared with existing privacy-preserving incentive mechanisms?

R: In PABP, devices bid to purchase FL model services. They perturb their bids locally before submission. The server uses these protected bids to select devices and allocate resources. It also calculates service fees without observing the original bids. The main differences are summarized below.

Gaussian bid perturbation. PABP protects the valuation of purchasing FL services before the bid reaches the server. The server then uses protected bids for both allocation and payment. BIT-FL [1] instead protects cost information through randomized worker selection and adds Gaussian noise to local models. PrivAim [2] protects participation costs and local models through differential privacy. RAFL [3] perturbs bids using the exponential mechanism and protects model aggregation with threshold Paillier encryption. In PABP, each device adds Gaussian noise to its bid locally before submitting the perturbed bid to the server. The original bid is kept locally. The server computes allocations and payments from the perturbed bids.

Myerson-Calibrated Payment. The calibration principle is straightforward: we introduce a payment correction term to reduce the difference between the executable payment and the ideal expected Myerson payment. The correction is calibrated before actual bidding, and the resulting payment is computed using protected bids. This differs from BIT-FL [1], which pays selected trainers and honest validators using role-dependent formulas based on cost bids and selection probabilities. PrivAim [2] uses reverse-auction payments to compensate data owners for FL participation. RAFL [3] pays sensing participants through a reverse auction that minimizes total platform payments under a budget constraint. In PABP, calibration instead adjusts the service fees paid by devices purchasing FL services. Under the revised calibration constraints, we have proved the uniform payment-error bound as δi=sup⁡b∈BN∣Eη[πi(b~)∣b]−πiref(bi;b−i)∣≤αiσi2π.{\delta }_{i}=\sup_{b\in {B}^{N}} \left\lvert {E}_{\eta }\left[ {\pi }_{i}(\tilde{b})\mid b\right]-{\pi }_{i}^{ref}({b}_{i};{b}_{-i})\right\rvert\le \frac{{\alpha }_{i}{\sigma }_{i}}{\sqrt{2\pi }}. Here, δᵢ denotes the conditional expected payment error, but we have proved an upper bound on this error. This bound also quantifies the approximate incentive compatibility (IC) upper bound 2δi2\delta_i and the approximate individual rationality (IR) upper bound δi\delta_i. Then we directly compared the metrics through a controlled experiment.

Table 4. Mean validation payment errors before and after calibration at two noise scales. Errors are normalized by αiBbid\alpha_iB^{bid} before averaging.

Noise scaleUncalibrated payment errorCalibrated payment errorRelative reduction
0.122.904%2.231%23.16%
0.307.423%6.179%16.76%

Calibration reduces mean validation payment error by 23.16% at σ=0.12\sigma=0.12 and 16.76% at σ=0.30\sigma=0.30. These results show that calibration brings the expected payment closer to the reference payment on the validation bids.

Joint resource allocation. PABP determines which devices participate and CPU resource and uploading powers under resource-limit, and deadline constraints. This differs from PrivAim [2], which recruits data owners with high-quality data and low participation costs. BIT-FL [1] selects workers probabilistically based on cost bids. RAFL [3] selects high-quality participants for sensing tasks while minimizing platform payments under a budget constraint. We first solve the independent minimum-cost resource optimization for each device to obtain its corresponding CPU frequency and upload power. We then compute the selection score based on this minimum resource cost, and adopt branch-and-bound to solve the device selection problem.

[1] Ying et al. (2025). BIT-FL: Blockchain-Enabled Incentivized and Secure Federated Learning Framework. IEEE Transactions on Mobile Computing, 24(2):1212–1229.

[2] Wang et al. (2023). PrivAim: A Dual-Privacy Preserving and Quality-Aware Incentive Mechanism for Federated Learning. IEEE Transactions on Computers, 72(7):1913–1927.

[3] Sun et al. (2026). RAFL: A Reverse Auction Federated Learning Framework with Non-Independent and Identically Distributed Data for Mobile Crowdsensing. Computer Networks, 283:112314.

Comments on the paper

5 comments
AuthorsOpenReview ↗

Global response (1/4)

We would like to express our sincere gratitude to all reviewers, CStl, DjLa, and DXRn, for dedicating their time and expertise to evaluate our paper. Their insightful comments are helpful for the improvement of our paper. In this global response, we will address the common concerns raised by the reviewers, including uniform payment-calibration error guarantees, zero-valuation devices selected under noise, dynamic-bid privacy, and differences in the comparison baselines, and further illustrate how each has been thoughtfully considered and incorporated into our revised draft and appendix.

  1. The primary concern raised by the reviewers CStl and DjLa is the lack of uniform payment calibration error guarantees and the limitations of finite-sample checks for establishing worst-case IC/IR bounds over the full bid space.

Thank you for highlighting the lack of uniform error guarantees. We agree that a uniform bound is needed to establish the approximate IC/IR guarantees of the payment rule. Following your suggestion, we have added an explicit payment-error bound in Section 4.4, with its proof in Appendix B.4. Specifically, under the revised payment-calibration constraints in Section 4.3 and details in Appendix A.5, we can prove that δi=sup⁡b∈BN∣Eη[πi(b~;θi∗)∣b]−πiref(bi;b−i)∣≤αiσi2π,\delta_i = \sup_{b \in B^N} \left| E_{\eta}\left[\pi_i(\tilde{b};\theta_i^{*}) \mid b\right] - \pi_i^{\mathrm{ref}}(b_i;b_{-i}) \right| \leq \frac{\alpha_i \sigma_i}{\sqrt{2\pi}}, where αi\alpha_i is the model-quality contribution coefficient of device ii, and σi\sigma_i is the noise scale. This bound holds for all bids, including bid profiles not sampled during calibration. Therefore, the approximate expected IC and IR guarantees in Section 4.4 (i.e., Theorem 1) give the following bounds.

The maximum expected IC is bounded by

εIC,i=2δi=2π αiσi.\color{black}{\varepsilon_{IC,i}=2\delta_i=\sqrt{\frac{2}{\pi}}\,\alpha_i\sigma_i.}

The expected IR is bounded by

−εIR,i=−δi=−αiσi2π.\color{black}{-\varepsilon_{IR,i}=-\delta_i=-\frac{\alpha_i\sigma_i}{\sqrt{2\pi}}.}

Moreover, in the paper, we state that calibration profiles are sampled from a finite bid space BN,{B}^{N}, where B=[0,Bbid]B=[0,{B}^{bid}] is defined in Section 3.1. This is because from the perspective of auction theory, each single device's bid should be within a scope [1] [2] [3]. In particular, [2] (Sections 2 and 5) and [3] (Sections 4.2 and 5.1) also consider settings with valuations and reports restricted to [0,1]. With this, the experimental results in Figures 4 and 5 in the paper just aim to validate our theoretical conclusions that the devised payment can satisfy the approximate IC/IR guarantees, i.e., Definitions 1 and 2. In the experiment evaluations, we set the bid domain B=[0,Bbid]B=[0,B^{bid}], with Bbid=1B^{bid}=1. The calibration profiles are sampled from BNB^N, and the reported bids are varied across the entire interval to evaluate the expected utility gain from misreporting, which follows the evaluation approach in [1] (Section 6.1). We also evaluate truthful expected utility against the approximate-IR lower bound.

[1] Li et al. (2024). Performance-Based Pricing for Federated Learning via Auction. PVLDB, 17(6):1269–1282.

[2] Dütting et al. (2019). Optimal Auctions through Deep Learning. ICML, PMLR 97:1706–1715.

[3] Rahme, Jelassi, and Weinberg (2021). Auction Learning as a Two-Player Game. ICLR.

  1. The second common concern, raised by the reviewers CStl and DjLa, is that bid noise may cause zero-valuation devices to be selected and incur resource costs.

We clarify this from two aspects. Firstly, a zero-valuation device after bid perturbation often has a lower probability of winning, as our considered selection criterion is to let devices with large valuations win; Secondly, a device with zero valuation means that it has no interest in participating in the model training process. Even if it is forced to submit a bid truthfully and wins because of its noisy bid, from the perspective of individual rationality, it would not voluntarily pay for winning this model it does not prefer. For example, in reality, a customer would not purchase an unwanted product. Similarly, selected devices with zero valuation are unwilling to pay in this case.

AuthorsOpenReview ↗

Global response (2/4)

  1. The third common concern, raised by the reviewers CStl and DjLa, is whether the bid privacy protections are still valid under multiple biddings, or the leakages of both allocations and payments.

Thank the reviewers for pointing out this issue. We agree that the original experiment considered a fixed true bid and did not adequately evaluate whether bid privacy protection is still valid under multiple biddings, where a device submits a new bid with fresh noise in each ordinary round. To this end, in the revised manuscript, we have established an explicit upper bound on the cumulative privacy budget under multiple biddings (see Appendix A.6.1).

ρi(T)≤∑ℓ∈OLΔb22(σi(ℓ))2=TΔb22σi2whenσi(ℓ)=σi.{\rho }_{i}^{(T)}\le \sum _{\ell \in {O}_{L}} \frac{{\Delta }_{b}^{2}}{2\bigl({\sigma }_{i}^{(\ell )}\bigr)^{2}}=\frac{T{\Delta }_{b}^{2}}{2{\sigma }_{i}^{2}}\quad \text{when}\quad { \sigma }_{i}^{(\ell )}={\sigma }_{i}.

Here, Δb\Delta_b is the bid sensitivity, which is the maximum permitted difference between neighboring true bids in each ordinary round. σi\sigma_i is the noise scale for device ii, namely the standard deviation of the Gaussian noise. T=∣OL∣T=\lvert O_L\rvert is the number of ordinary rounds in which device ii submits a new perturbed bid. The privacy budget ρi(T)\rho_i^{(T)} characterizes the extent to which the server can distinguish between two neighboring true-bid sequences after observing allocations and payments, thereby reflecting its ability to infer true bids. A smaller privacy budget provides a stronger privacy guarantee. We establish an upper bound on the cumulative privacy budget across multiple rounds. This bound shows that the server’s ability to infer true bids is limited.

Furthermore, since allocations and payments are obtained from perturbed bids, devices' allocation and payment information therefore also falls within the scope of privacy protection, and there is no cumulative leakage when bids change or when the server launches attacks by exploiting allocation and payment information.

For better illustrations, we also add an experiment to numerically evaluate whether bid privacy protection is still valid under multiple biddings.

Dynamic-bid privacy protection experiment. Each device draws a new valuation independently from a uniform distribution on [0,1] in every ordinary round, and perturbs it with fresh Gaussian noise of standard deviation 0.12. The server knows the bid distribution and perturbation mechanism, and uses the available history to reconstruct the device’s true bid.

Table 1. Current true-bid reconstruction error (RMSE) after 80 ordinary rounds: PABP versus the auction without bid privacy (Appendix C.1).

ConfigurationPABP (RMSE)Auction without bid privacy (RMSE)
MNIST0.105268 ± 0.0030400
Fashion-MNIST0.105654 ± 0.0028310
CIFAR-100.104622 ± 0.0035690

The root mean square error (RMSE) measures the discrepancy between reconstructed and original bids. A value of zero means exact bid recovery. A positive value indicates imperfect recovery and some privacy protection against the evaluated attack. PABP’s mean RMSE remains 0.1041–0.1094 across 1–80 ordinary-round releases, compared with zero for the Non-private auction, demonstrating that the proposed PABP can protect the recovery of current true bids under dynamic bidding.

AuthorsOpenReview ↗

Global response (3/4)

  1. The fourth common concern, raised by the reviewers CStl, DjLa, and DXRn, is that the baseline settings differ from PABP in their privacy protection and incentive settings, which affects the fairness of the comparisons.

Thank the reviewers for pointing out the differences between the baseline settings and ours. We first compare PABP with the three related studies recommended by reviewer DXRn: BIT-FL [1], PrivAim [2], and RAFL [3].

BIT-FL [1] protects cost information through randomized worker selection and protects local models. Our PABP instead directly protects valuation bids before submission. PrivAim [2] and RAFL [3] also perturb participants’ bids bb for privacy protection. Therefore, we add experiments on the ImageNet dataset to compare PABP with PrivAim-bid-adapted and RAFL-bid-adapted.

Our comparison metric is the social welfare ratio, defined as the ratio of the aggregated expected true social welfare to the full-information optimal social welfare without bid perturbation. This ratio measures how much of the optimal social welfare each method retains while protecting bids. A higher ratio indicates better welfare preservation and less social welfare loss.

Table 2. Social welfare ratios of PABP, PrivAim-bid-adapted, and RAFL-bid-adapted.

MethodSocial welfare ratio
PABP97.215%
PrivAim-bid-adapted88.129%
RAFL-bid-adapted87.560%

Table 2 shows that PABP achieves a higher social welfare ratio and less social welfare loss than both adapted baselines. Based on our analysis, we believe that PABP’s additional resource optimization contributes to this improvement. PABP jointly considers winner selection and resource allocation. It computes optimal CPU frequencies and upload powers and assigns these configurations to the selected devices, reducing their resource costs. PrivAim and RAFL do not optimize CPU frequencies or upload powers.

[1] Ying et al. (2025). BIT-FL: Blockchain-Enabled Incentivized and Secure Federated Learning Framework. IEEE Transactions on Mobile Computing, 24(2):1212–1229. Paper.

[2] Wang et al. (2023). PrivAim: A Dual-Privacy Preserving and Quality-Aware Incentive Mechanism for Federated Learning. IEEE Transactions on Computers, 72(7):1913–1927. DOI.

[3] Sun et al. (2026). RAFL: A Reverse Auction Federated Learning Framework with Non-Independent and Identically Distributed Data for Mobile Crowdsensing. Computer Networks, 283:112314. DOI.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.