acceptodds
Under review as a conference paper at ICLR 2027

GatedAFL: Privacy-Preserving and Resource-Efficient Asynchronous Personalized Federated Learning with Trainable Thresholds

Abstract

Conventional federated learning (FL) methods typically rely on exchanging model parameters or loss gradients between clients and the server, which incurs substantial communication overhead and exposes sensitive information embedded in model updates. Threshold-gated training offers a more compact alternative, in which a learnable threshold per neuron determines whether its associated parameters are retained or pruned, and only these thresholds need to be communicated. This paper proposes GatedAFL, a threshold-gating method for asynchronous personalized FL. Clients keep their personalized parameters strictly local and upload only the threshold increment accumulated over their local updates. We first prove mathematically that threshold sharing never leaks more information than gradient sharing, and then quantify the privacy gap between the two approaches. Extensive experiments confirm that GatedAFL achieves enhanced resilience against privacy attacks while maintaining competitive accuracy. It reduces communication overhead by two to three orders of magnitude compared to FedAsync, and lowers computational cost by relative to SpaFL.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
asynchronous federated learning, pruning, privacy preservation, personalized federated learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_WMDhOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a new federated learning algorithm by only sharing learned thresholds instead of parameters/gradients across clients, which is supposed to reduce communication overhead.

Critical strengths

The idea of using learnable threshold is an interesting alternative to typical FL algorithms.

Critical weaknesses
  1. The paper mentions better privacy protection via threshold sharing but the paper does not have formal analysis of privacy guarantee rather than heuristics.

  2. The threat model of the claim is very unclear.

  3. The experiments are very limited to small set of image datasets

Questions for authors

Does the attacker in your theory have access to model weights? It appears not but the theory assumes so?

2 comments

W1: Formal privacy analysis

We would like to argue that our privacy claims are supported by a formal information-theoretic analysis, rather than by heuristics alone. Specifically, Theorem 1 establishes that, for the same model parameters, a single-step threshold gradient reveals no more information about the private data than the corresponding parameter gradient, as measured by mutual information. Theorem 2 and Appendix A.2 further quantify the information-leakage difference under their stated assumptions.

Mutual information is a widely used privacy measure. For example, Calmon and Fawaz (2012) use it to characterize privacy against statistical inference, and Elkordy et al. (2023) use it to analyze privacy leakage in federated learning with secure aggregation. Our results therefore provide a theoretical comparison between the two sharing mechanisms, in addition to the empirical reconstruction evaluation.

Manuscript revision. Section 4 now clarifies the information-theoretic basis and scope of our privacy analysis, including its relationship to optimal reconstruction risk and relevant prior work.

W2/Question: Attacker information in the theory and protocol

Thank you for the question.

Yes. In our theoretical comparison, the attacker is assumed to know the model parameters ww. We provide the same ww as side information when evaluating the information revealed by the threshold gradient and the parameter gradient. In the reconstruction experiments, the attacker is also given the initial local weights wk(t)\boldsymbol{w}_k(t), the initial thresholds τ(t)\boldsymbol{\tau}^{(t)}, and the true labels, and replays the local training steps to match the observed update. We consider known ww for two main reasons.

First, our gradient-matching baselines, IG (Geiping et al., 2020) and GradInversion (Yin et al., 2021), assume that ww is available as side information. Reconstructing images using only GG without ww is generally much more difficult, and we are not aware of a general method that achieves this without further assumptions. We therefore also provide ww to align our analysis with these baselines. Methods that do not require the same weight information either exploit additional structure or recover more limited information. For example, IG's Appendix D gives a single-sample analytical recovery result using the specific structure of a biased fully connected layer, while BlindSage (Arazzi et al., 2024) recovers only data labels without knowing the server model in vertical federated GNNs. Therefore, to guarantee a relatively general and strong attack, we assume that the attacker has access to the model weights.

Second, we use the same WW to generate the parameter gradient GG and threshold gradient ZZ to ensure a fair comparison. If the weights differ, the model state itself may affect the privacy gap. In particular, different weights change the activations and derivatives through which private data affect the gradients. Comparing gradients generated at different weights could therefore confound the effect of the sharing mechanism with the effect of the model state.

Manuscript revision. Section 4 now states explicitly that ww is theoretical side information, not a transmitted message. It explains the alignment with IG/GradInversion and defines both gradients at the same model state for a fair comparison.

References for this response

F. du P. Calmon and N. Fawaz. Privacy Against Statistical Inference. Allerton, 2012.

A. R. Elkordy et al. How Much Privacy Does Federated Learning with Secure Aggregation Guarantee? Proceedings on Privacy Enhancing Technologies, 2023(1):510-526, 2023.

J. Geiping, H. Bauermeister, H. Dr"{o}ge, and M. Moeller. Inverting Gradients: How Easy Is It to Break Privacy in Federated Learning? NeurIPS, 2020.

H. Yin et al. See through Gradients: Image Batch Recovery via GradInversion. CVPR, 2021.

W3: Dataset coverage

Thank you for your suggestion to broaden the experimental evaluation.

We believe that the three datasets are well established and appropriate benchmarks for evaluating GatedAFL. Notably, TCT (Yu et al., 2022), SpaFL (Kim et al., 2024), and SkyMask (Yan et al., 2024) all use the same three datasets: FMNIST, CIFAR-10, and CIFAR-100. These studies address federated optimization, model sparsification, and security, which are closely related to the objectives of our work.

Furthermore, our experiments cover different classification complexities, ranging from 10 to 100 classes, and diverse model architectures, including LeNet, ViT, and ResNet-18. All experiments involve 100 clients with heterogeneous data distributions, allowing us to evaluate accuracy, model sparsity, communication efficiency, and resistance to reconstruction attacks across different datasets and model architectures.

Therefore, our benchmark selection aligns with established research practices and provides a comprehensive evaluation of GatedAFL across multiple experimental settings.

References for this response

Y. Yu, A. Wei, S. P. Karimireddy, Y. Ma, and M. I. Jordan. TCT: Convexifying Federated Learning using Bootstrapped Neural Tangent Kernels. NeurIPS, 35, 2022.

M. Kim, W. Saad, M. Debbah, and C. S. Hong. SpaFL: Communication-Efficient Federated Learning with Sparse Models and Low Computational Overhead. NeurIPS, 37:86500--86527, 2024.

P. Yan, H. Wang, T. Song, Y. Hua, R. Ma, N. Hu, M. R. Haghighat, and H. Guan. SkyMask: Attack-Agnostic Robust Federated Learning with Fine-Grained Learnable Masks. ECCV, pp. 291--308, 2024.

Reviewer_ekYWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper builds on SpaFL and adapts it to asynchronous personalized federated learning. Each client keeps its model weights local and uploads the change in its thresholds after local training. The server collects these updates in a buffer and adds their average to the current global thresholds. The authors also analyze privacy at fixed model weights and derive a leakage ratio under Gaussian and isotropy assumptions. Experiments on CIFAR-10, CIFAR-100, and FMNIST report accuracy close to FedAsync and SpaFL, with lower communication costs than FedAsync and fewer FLOPs than SpaFL. They also show poorer image reconstruction from threshold information under IG and GradInversion.

Critical strengths
  1. The protocol extends threshold sharing to clients that train and reconnect at different times while keeping model weights local. Its increment rule avoids directly replacing the server's current thresholds with stale absolute values.

  2. At fixed weights, the analysis expresses the privacy gap as data information present in the full gradient but absent from the threshold gradient. Appendix A.3 also clarifies that sending increments provides no additional privacy over sending final thresholds when the server knows the starting threshold.

Critical weaknesses
  1. Section 5.3 does not explain how the attacks match threshold updates or whether they know the local weights, and tests only single-sample, single-step observations. Privacy gains for the actual cumulative uploads remain uncertain.

  2. The main comparison uses a single delay setting, and Appendix B does not compare the aggregation rules under controlled changes in staleness. This limits the evidence for the claimed staleness tolerance.

  3. The test-set partitioning and averaging across clients are unspecified, so it is unclear whether all methods are evaluated on the same target distribution.

Questions for authors
  1. Can you report a reconstruction attack adapted to the actual EE-step threshold increment, specifying access to local weights and the optimization budget?
  2. How do GatedAFL and asynchronous SpaFL compare as staleness increases, with the number of completed client updates held constant?
  3. How are client test sets constructed and their accuracies averaged, and do all methods use this same protocol?
3 comments

W1/Q1: Attacker information and privacy of cumulative uploads(1/2)

Privacy of Single-Step and Multi-Step Updates

Thank you for your question. Indeed, our main theoretical results in Theorems 1 and 2 and Appendix A.2 consider only single-step threshold and parameter gradients. Analyzing the privacy of multi-step updates is challenging because each local gradient depends on the model state produced by earlier steps, while those intermediate states are hidden from the observer when only the accumulated update is shared.

Existing theoretical analyses of multi-step gradient privacy often require strong assumptions. For example, Zhu et al. (2023) analyze the directional alignment between a surrogate gradient and a cumulative update under smoothness, strong convexity, near-two-dimensional gradient concentration, and additional alignment, step-size, and noise conditions. In a different setting, Annamalai (2024) gives nonconvex DP-SGD counterexamples showing that hiding intermediate states does not universally guarantee additional privacy amplification. To our knowledge, an assumption-free characterization of privacy leakage from cumulative parameter gradients in general settings remains unavailable, and threshold updates face the same difficulty. Establishing broader conditions for threshold privacy gains and testing them against attacks on the transmitted increments remain important future work.

Nevertheless, a comparison is possible under additional conditions. We have added Appendix A.4 to compare the privacy leakage of the multi-step threshold increment and the multi-step parameter update. Specifically, we identify a sufficient condition: given the parameter update, the common initial model state, and the training settings, the gradient-energy and regularization terms accumulated over the local steps reveal no additional information about the private data. Under this condition, sharing the threshold increment leaks no more information than sharing the parameter update. The privacy gain is strict if the parameter update contains private-data information that is not retained by the threshold increment. These results provide insight into multi-step threshold privacy, although the sufficient condition remains to be evaluated in practical training. Establishing broader conditions for a privacy gain and assessing whether they hold in practice are promising directions for future work.

Attacker Knowledge of Model Weights

Regarding whether the attacker knows the local weights, as stated in our paper, we provide both threshold gradient and parameter gradient attackers with the same ww as side information in the theoretical proof. We explain the reasons for this assumption below, but in actual communication, GatedAFL does not transmit ww, but only threshold increments. Therefore, the attacker we assume in the theory is a relatively strong attacker. For the same observed message, attacking without ww is no easier than attacking with the side information provided in our analysis. There are two main reasons for considering known ww in the proof.

First, our gradient-matching baselines, IG (Geiping et al., 2020) and GradInversion (Yin et al., 2021), assume that ww is available as side information. Reconstructing images using only GG without ww is generally much more difficult, and we are not aware of a general method that achieves this without further assumptions. We therefore also provide ww to align our analysis with these baselines. Methods that do not require the same weight information either exploit additional structure or recover more limited information. For example, IG's Appendix D gives a single-sample analytical recovery result using the specific structure of a biased fully connected layer, while BlindSage (Arazzi et al., 2024) recovers only data labels without knowing the server model in vertical federated GNNs. Therefore, to guarantee a relatively general and strong attack, we assume that the attacker has access to the model weights.

Second, we use the same WW to generate the parameter gradient GG and threshold gradient ZZ to ensure a fair comparison. If the weights differ, the model state itself may affect the privacy gap. In particular, different weights change the activations and derivatives through which private data affect the gradients. Comparing gradients generated at different weights could therefore confound the effect of the sharing mechanism with the effect of the model state.

Moreover, GG and ZZ represent alternative sharing choices; they are not transmitted together in GatedAFL. Comparing GG and ZZ generated at different weights therefore has limited practical relevance to this choice.

W1/Q1: Attacker information and privacy of cumulative uploads(2/2)

We evaluate reconstruction attacks on the actual cumulative threshold increment Δτk\Delta\boldsymbol{\tau}_k transmitted by GatedAFL after 10 local training steps. The attacker is given the initial local weights wk(t)\boldsymbol{w}_k(t), thresholds τ(t)\boldsymbol{\tau}^{(t)}, and true labels, and replays all 10 steps to match the observed increment. Both IG and GradInversion use 1200 optimization iterations with a learning rate of 0.05, one restart. IG uses one candidate reconstruction, whereas GradInversion uses four randomly initialized candidates. For each attack, identical optimization budgets are used for the threshold increment and gradient sharing settings. The reconstruction results are reported below:

AttackMetricCIFAR-10CIFAR-100FMNIST
ThresholdGradientThresholdGradientThresholdGradient
IGMSE0.1510.0100.2080.0130.2600.003
PSNR8.21720.0416.82218.7555.85024.780
SSIM0.0360.7490.0080.5650.0750.922
GradInversionMSE0.1430.0480.2220.0210.2490.004
PSNR8.45713.1886.54416.8416.03524.461
SSIM0.0490.4020.0180.4680.0130.902

Across all three datasets and both attacks, reconstructions from the transmitted threshold increments consistently exhibit higher MSE and lower PSNR and SSIM than those from model gradients, demonstrating substantially stronger resistance to reconstruction attacks. Complete experimental results are provided in Appendix D.2.

Manuscript revision. We have strengthened our privacy analysis through theoretical clarification and additional experimental evaluation. Appendix A.4 establishes conditions under which cumulative threshold increments leak no more information than cumulative parameter updates. Section 4 clarifies the threat model by granting both attackers the same local weights ww, consistent with IG and GradInversion, while explicitly stating that GatedAFL does not transmit these weights. Complementing the theoretical analysis, Appendix D.2 provides a reconstruction evaluation of the actual cumulative threshold increments transmitted by GatedAFL.

References for this response

M. Arazzi et al. Label Inference Attacks against Node-level Vertical Federated GNNs. arXiv:2308.02465, April 2024 version.

J. Geiping, H. Bauermeister, H. Dr"{o}ge, and M. Moeller. Inverting Gradients: How Easy Is It to Break Privacy in Federated Learning? NeurIPS, 2020.

H. Yin et al. See through Gradients: Image Batch Recovery via GradInversion. CVPR, 2021.

J. Zhu, R. Yao, and M. B. Blaschko. Surrogate Model Extension (SME): A Fast and Accurate Weight Update Attack on Federated Learning. ICML, PMLR 202:43228-43257, 2023.

M. S. M. S. Annamalai. It's Our Loss: No Privacy Amplification for Hidden State DP-SGD With Non-Convex Loss. AISec, 2024.

W2/Q2: Comparison under controlled staleness

Thank you for the question.

To evaluate robustness to staleness, we compare the incremental threshold aggregation of GatedAFL with the threshold averaging of SpaFL under controlled staleness levels. Both methods use the same buffer size and matched server aggregation counts, ensuring an equal number of completed client updates. The results at s=8s=8 are shown below:

DataGatedAFLSpaFL
Acc. (%)DensityAcc. (%)Density
FMNIST90.310.38989.960.533
CIFAR-1071.940.50872.030.656
CIFAR-10045.130.93245.680.998

GatedAFL consistently maintains lower model density while achieving competitive accuracy across all three datasets. On CIFAR-10, this density advantage persists across all tested staleness levels, demonstrating the effectiveness of incremental threshold aggregation in preserving model sparsity under stale updates. Complete experimental results are provided in Appendix B.5.

Manuscript revision. We have added the controlled comparison with asynchronous SpaFL across staleness levels in Appendix B.5, including the complete accuracy and model-density results in Table 3.

W3/Q3: Test-set construction and averaging

Thank you for the question.

Each client is evaluated on a held-out test partition that follows its local training label distribution. We report the average accuracy across K=100K=100 clients, with equal weight assigned to each client. All methods are evaluated using the same client-specific test partitions and averaging procedure, ensuring a fair comparison under identical evaluation conditions.

Reviewer_eaXqOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GatedAFL, an asynchronous personalized federated learning method that shares only threshold updates instead of model weights. It studies privacy leakage and shows lower communication and computation costs while keeping competitive accuracy.

Critical strengths
  1. Important problem. The paper addresses communication, computation, and privacy challenges in federated learning.
  2. Simple design. The method is easy to understand and supports asynchronous training.
  3. Broad experiments. The paper tests different datasets and model architectures. It also studies the effects of several key settings.
Critical weaknesses
  1. Limited novelty over SpaFL. GatedAFL largely builds on SpaFL, and the two methods show nearly identical accuracy and communication costs. The main algorithmic difference appears to be incremental threshold aggregation, but its benefit is not clearly demonstrated.

  2. Restrictive assumptions in Theorem 2. Theorem 2 assumes that both the marginal and data-conditioned gradient covariance matrices are isotropic and full-rank. These assumptions are questionable in high-dimensional settings, especially when the number of local training samples nn is much smaller than the gradient dimension dd.

  3. Privacy evaluation does not match the actual protocol. The privacy experiments attack single-step threshold gradients, whereas GatedAFL transmits threshold increments accumulated over multiple local steps. Thus, the experiments do not directly establish the privacy of the transmitted information.

  4. Unclear evaluation under data heterogeneity. Figure 8 shows higher accuracy with greater data heterogeneity, but the test distribution is not clearly specified. Without a local-only baseline, it is also unclear whether sharing thresholds provides meaningful improvements over independent local training.

Questions for authors
  1. How is GatedAFL different from SpaFL? What improvement does incremental threshold aggregation provide over SpaFL, particularly under stale updates?

  2. Under what conditions can these covariance assumptions hold when n≪dn\ll d, and are they satisfied in the experiments?

  3. How well does GatedAFL resist reconstruction attacks targeting the actual transmitted threshold increments?

  4. Is accuracy evaluated on a common global test set or label-skewed local test sets, and how does GatedAFL compare with local-only training?

4 comments

W1/Q1: Relationship to SpaFL and increment aggregation

Thank you for your careful comparison with SpaFL and for asking us to clarify the contributions of our work.

We would first like to clarify that our contributions to SpaFL lie in two aspects: asynchronous aggregation and theoretical privacy analysis. In particular, we provide a rigorous mutual-information analysis of SpaFL's threshold mechanism and prove that threshold gradients computed from the same dataset and model parameters reveal no more information than parameter gradients.

Previously, threshold sharing in SpaFL was primarily a communication-efficient method. Our work is, to the best of our knowledge, the first to establish a theoretical privacy guarantee for this mechanism. This result provides a privacy rationale for threshold sharing even when communication resources are sufficient, and provides a stronger theoretical foundation for its use beyond communication efficiency.

Beyond the theoretical analysis, we assess whether existing attacks can exploit shared thresholds. Our evaluation shows that classical attacks designed for model parameters and parameter gradients do not readily transfer to thresholds. To the best of our knowledge, attacks specifically targeting threshold gradients remain unexplored, and designing them presents a distinct challenge. The results in Fig. 4 and Table I provide empirical evidence that threshold sharing may offer privacy advantages against the attacks considered.

Regarding asynchronous aggregation, our controlled staleness experiments further demonstrate the robustness of incremental threshold aggregation in GatedAFL compared with threshold averaging in SpaFL. At s=8s=8, GatedAFL maintains competitive accuracy while achieving lower model density across all three datasets. On CIFAR-10, GatedAFL consistently maintains lower model density across all tested staleness levels, highlighting its effectiveness in preserving model sparsity under stale updates. The results at s=8s=8 are shown below:

DataGatedAFLSpaFL
Acc. (%)DensityAcc. (%)Density
FMNIST90.310.38989.960.533
CIFAR-1071.940.50872.030.656
CIFAR-10045.130.93245.680.998

Complete experimental results are provided in Appendix B.5.

Manuscript revision. We have strengthened our comparison with SpaFL from both theoretical and experimental perspectives. In Section 4, we further clarify our theoretical contribution by showing that threshold sharing has an information-theoretic privacy guarantee in addition to its communication benefit. In Appendix B.5, we further compare our method with asynchronous SpaFL under different staleness levels and report the complete accuracy and model density results in Table 3.

W2/Q2: Scope of the covariance assumptions

Thank you for the question.

Theorem 2 and the intermediate result in Appendix A.2 build on Theorem 1 and introduce additional assumptions to obtain clearer and more explicit conclusions. Such assumptions have been used in related analyses. Specifically, Chechik et al. (2005) use a joint Gaussian model to analyze the information bottleneck; Mandt et al. (2017, Section 3.2) assume Gaussian gradient noise with a full-rank covariance in their SGD analysis; and Jastrzębski et al. (2018, Section 2.3) consider an isotropic gradient covariance. Our result in Appendix A.2 requires joint Gaussianity and full-rank marginal and data-conditioned gradient covariances, while Theorem 2 additionally assumes isotropy.

Whether the full-rank assumption holds when n≪dn\ll d depends on the source of gradient randomness. For example, a nondegenerate dd-dimensional Gaussian distribution has a full-rank population covariance even if only n≪dn\ll d samples are observed. Additional randomness from dropout or noise may also increase the rank of the gradient covariance. However, if the client’s dataset is fixed and gradient randomness comes only from sampling its nn per-example gradients, the covariance has rank at most n−1n-1 and is therefore singular when n≪dn\ll d. These examples do not by themselves establish joint Gaussianity or isotropy, whose validity in practice requires further investigation.

We evaluated MNIST with Gaussian noise added to the effective-weight gradients (σ≈0.447\sigma\approx0.447). Gaussianity was checked at 19 fixed model states, using batches from ten clients at each state. The isotropy value is from a 2D projection in the first prespecified confirmation run. Accuracy was compared across ten paired training runs with and without noise.

MeasureObserved value (criterion)
Gaussian-test statistic0.6160.616–0.9080.908 in 18/19 states, 1.0231.023 in the remaining state (≤1\leq 1: no significant departure from Gaussianity).
Relative deviation from isotropy0.00920.0092, or 0.920.92\\% (00: exact isotropy; smaller values indicate a closer approximation).
Test accuracy changeMean: +0.309+0.309 percentage points (required lower bound: >−0.5>-0.5 percentage points).

These results provide empirical support for the practical attainability of approximately Gaussian and nearly isotropic conditional gradient statistics through suitable noise injection, without sacrificing test accuracy.

Manuscript revision. We have moved part of Appendix A.2 into the main text. Unlike Theorem 2, this result does not require isotropy, and it still quantifies the mutual-information gain from threshold sharing. This change improves the flow of the theoretical analysis and introduces the assumptions more gradually, providing a better balance between generality and clarity.

References for this response

G. Chechik, A. Globerson, N. Tishby, and Y. Weiss. Information Bottleneck for Gaussian Variables. JMLR, 6:165-188, 2005.

S. Mandt, M. D. Hoffman, and D. M. Blei. Stochastic Gradient Descent as Approximate Bayesian Inference. JMLR, 18(134):1-35, 2017. Section 3.2.

S. Jastrzębski et al. Three Factors Influencing Minima in SGD. arXiv:1711.04623, September 2018 version. Section 2.3.

W3/Q3: Privacy of the actual multi-step threshold increment

Thank you for your question. Indeed, our main theoretical results in Theorems 1 and 2 and Appendix A.2 consider only single-step threshold and parameter gradients. Analyzing the privacy of multi-step updates is challenging because each local gradient depends on the model state produced by earlier steps, while those intermediate states are hidden from the observer when only the accumulated update is shared.

Existing theoretical analyses of multi-step gradient privacy often require strong assumptions. For example, Zhu et al. (2023) analyze the directional alignment between a surrogate gradient and a cumulative update under smoothness, strong convexity, near-two-dimensional gradient concentration, and additional alignment, step-size, and noise conditions. In a different setting, Annamalai (2024) gives nonconvex DP-SGD counterexamples showing that hiding intermediate states does not universally guarantee additional privacy amplification. To our knowledge, an assumption-free characterization of privacy leakage from cumulative parameter gradients in general settings remains unavailable, and threshold updates face the same difficulty. Establishing broader conditions for threshold privacy gains and testing them against attacks on the transmitted increments remain important future work.

Nevertheless, a comparison is possible under additional conditions. We have added Appendix A.4 to compare the privacy leakage of the multi-step threshold increment and the multi-step parameter update. Specifically, we identify a sufficient condition: given the parameter update, the common initial model state, and the training settings, the gradient-energy and regularization terms accumulated over the local steps reveal no additional information about the private data. Under this condition, sharing the threshold increment leaks no more information than sharing the parameter update. The privacy gain is strict if the parameter update contains private-data information that is not retained by the threshold increment. These results provide insight into multi-step threshold privacy, although the sufficient condition remains to be evaluated in practical training. Establishing broader conditions for a privacy gain and assessing whether they hold in practice are promising directions for future work.

Experimentally, we evaluate the actual cumulative threshold increment Δτk\Delta\tau_k transmitted by GatedAFL after 10 local training steps under both IG and GradInversion, using identical optimization budgets within each attack. Both IG and GradInversion use 12001200 optimization iterations with a learning rate of 0.050.05, one restart. IG uses one candidate reconstruction, whereas GradInversion uses four randomly initialized candidates. Across all three datasets, reconstructions from the transmitted increments consistently exhibit higher MSE and lower PSNR and SSIM than those from model gradients, demonstrating substantially stronger resistance to reconstruction attacks.

The reconstruction results are reported below:

AttackMetricCIFAR-10CIFAR-100FMNIST
ThresholdGradientThresholdGradientThresholdGradient
IGMSE0.1510.0100.2080.0130.2600.003
PSNR8.21720.0416.82218.7555.85024.780
SSIM0.0360.7490.0080.5650.0750.922
GradInversionMSE0.1430.0480.2220.0210.2490.004
PSNR8.45713.1886.54416.8416.03524.461
SSIM0.0490.4020.0180.4680.0130.902

Complete experimental results presented in Appendix D.2.

Manuscript revision. We have strengthened our privacy analysis of cumulative threshold increments from both theoretical and experimental perspectives. Appendix A.3 now clarifies the scope of its leakage bound, while the new Appendix A.4 establishes conditions under which multi-step threshold increments leak no more information than multi-step parameter updates, including a criterion for strict privacy gain. Complementing these theoretical results, Appendix D.2 provides a reconstruction evaluation of the actual multi-step threshold increments transmitted by GatedAFL.

References for this response

J. Zhu, R. Yao, and M. B. Blaschko. Surrogate Model Extension (SME): A Fast and Accurate Weight Update Attack on Federated Learning. ICML, PMLR 202:43228-43257, 2023.

M. S. M. S. Annamalai. It's Our Loss: No Privacy Amplification for Hidden State DP-SGD With Non-Convex Loss. AISec, 2024.

W4/Q4: Evaluation distribution and the local-only baseline

Thank you for the question.

We report the average accuracy across K=100K=100 clients, each evaluated on a held-out test partition following its local training label distribution. As α\alpha decreases, client data become increasingly concentrated in fewer classes. Since the test partitions preserve the corresponding label skew and the models remain personalized, stronger data heterogeneity can favor client specialization, explaining the trend observed in Figure 8. To further evaluate the benefit of threshold exchange, we compare GatedAFL with local training without threshold exchange:

GatedAFLLocal-only
DataAcc. (%)DensityAcc. (%)Density
FMNIST89.970.37183.750.759
CIFAR-1069.420.36272.670.770
CIFAR-10045.240.91745.970.990

Compared with local training, GatedAFL substantially reduces model density across all three datasets, particularly on FMNIST and CIFAR-10, while maintaining competitive personalized accuracy.

Manuscript revision. We have added a comparison with local-only training in Section 5.2 to evaluate the benefit of threshold exchange in terms of personalized accuracy and model density.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.