acceptodds
Under review as a conference paper at ICLR 2027

CloakFL: Mitigating Global Model Exposure in Federated Learning via Convergence-Aware Proxy Distribution

Abstract

In federated learning, the server repeatedly sends the global model to participating clients for local training, exposing the model to the risk of leakage or redistribution outside the federation. Proxy Model Distribution (PMD) mitigates this risk by distributing intentionally degraded proxy models instead of the clean global model. Existing PMD methods, however, face a protection–convergence trade-off: stronger degradation reduces the proxy’s inference performance but impedes global convergence, whereas weaker degradation better preserves global convergence but provides weaker protection. We propose CloakFL, a PMD method that adds convergence-aware perturbations to the global model. CloakFL constructs perturbation directions derived from the previous round’s local training, enabling effective degradation with relatively small perturbations. It then scales each perturbation according to the target client’s latest update to better preserve global convergence. Through convergence analysis and experiments on CNN and Transformer architectures, we show that the performance of the proxy models remains substantially below that of the global model, while keeping the global model’s performance close to that of standard federated learning.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Model Protection, Proxy Model, Client Drift

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_59yLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies global-model exposure in federated learning through proxy model distribution. CloakFL constructs client-specific perturbations using centered local updates from the previous round and scales them layer-wise according to the receiving client’s latest update magnitude. The goal is to degrade the distributed proxies while preserving the utility of server aggregation. The paper provides a convergence bound for a generalized perturbed FedAvg process under strong convexity, full participation, and additional perturbation assumptions. Experiments cover four vision and language tasks, cross-silo and cross-device settings, and single-client and bounded multi-client recovery. The reported results show larger server–proxy performance gaps than FedQSN at comparable server performance.

Critical strengths

1.The perturbation design uses a meaningful optimization signal. Constructing proxies from client drift, rather than arbitrary noise, provides a plausible way to degrade individual models while limiting disruption to aggregation. The direction and magnitude ablations offer empirical support for this approach.

2.The evaluation extends beyond raw proxy degradation. The paper includes vision and language architectures, different heterogeneity levels, and recovery through local training, fine-tuning, and coalition averaging. These comparisons are useful because a low-performing raw proxy alone does not establish protection against a client that can continue training it.

Critical weaknesses

1.The convergence assumptions are not established for the proposed algorithm. Assumption 3.3 requires the aggregated perturbation to be conditionally unbiased, and Appendix F uses this requirement to eliminate a cross term. Centering the original client drifts does not guarantee this property after source-direction normalization and client-specific layer-wise scaling. For example, with two clients having drifts (v) and (-v), the exclusion of self-assignment forces a swap. If their reference update norms are (a) and (b), the average perturbation is proportional to ((b-a)v), which is generally nonzero. The algorithm also does not establish the learning-rate-scaled moment bounds in Assumptions 3.4–3.5. Although the appendix acknowledges a theory–practice gap, the current theorem therefore provides a conditional result for a generalized perturbation process rather than a demonstrated convergence guarantee for Algorithm 1.

2.The recovery evaluation is too limited to substantiate the protection claim. Recovery fine-tuning uses only one local epoch at a fixed reduced learning rate, and the evaluation does not test stronger optimization or exploitation of models collected across rounds. Substantial recovery already occurs under this limited protocol: in the cross-device E2E setting with (\beta=0.1), CloakFL’s raw proxy BLEU is 18.5, but single-client recovery reaches 57.1 against a server score of 62.9. Furthermore, the reported “Max” is the maximum of attack-specific mean scores, and coalition results average ten sampled groups. These statistics do not capture whether one client or a favorable coalition can recover a valuable model, which is directly relevant to model leakage.

Questions for authors

1.Does Algorithm 1 satisfy conditional unbiasedness after layer-wise normalization, client-specific scaling, and exclusion of self-assignment? If so, please provide the argument. How are the moment bounds in Assumptions 3.4–3.5 established for stored historical updates?

2.Theorem 3.1 bounds the average initial proxy model (\bar w_T^{(0)}). What corresponding guarantee follows for the clean server model (w_T) returned by Algorithm 1?

  1. Does CloakFL retain its advantage when recovery uses multiple epochs, learning-rate selection, and model histories from previous rounds, while retaining the same restriction on available task data?

  2. What are the best-client and best-coalition recovery results, and the distribution of recovery performance, rather than only the maximum of attack-specific means?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 59yL (1/5)

We sincerely thank the reviewer for the constructive feedback. To address the concerns regarding the identified weaknesses, we provide the following clarifications together with additional experimental results.

W1. The convergence assumptions are not established for the proposed algorithm.

In CloakFL without scaling, the mean of the client drifts assigned in a round is exactly zero, so the mean of the aggregated perturbation is also exactly zero. Therefore, this configuration satisfies the conditional unbiasedness assumption of Theorem 3.1. In addition, because the average of the proxy models is exactly equal to the clean server model without scaling, the convergence bound in Theorem 3.1 also applies to the server model in this configuration.

On Tiny-ImageNet, server performance without scaling was similar to that with scaling, and the performance of both the proxies and the recovered models was below that of the server. Client-specific scaling was added to improve protection effectiveness in settings such as 20 Newsgroups and to control the trade-off between protection effectiveness and server performance. Exact cancellation is not guaranteed with scaling, but the aggregated perturbation remains close to zero, as shown in Figure 1(b) and the table below.

Tiny-ImageNet

Individual perturbation normAggregated perturbation norm
3.79320.3675

20 Newsgroups

Individual perturbation normAggregated perturbation norm
6.52970.7580

The entries in the table below report the maximum / minimum values across proxy models or recovery sets.

Tiny-ImageNet

ScalingServerProxyProxyTrainProxyFTColl.ProxyColl.TrainColl.FT
Unscaled29.368.51 / 0.5310.22 / 4.7917.58 / 4.6320.08 / 10.7523.88 / 19.0726.03 / 18.79
Scaled29.6910.61 / 4.4612.20 / 1.3518.82 / 11.1323.47 / 15.9523.27 / 18.6427.62 / 22.71

20 Newsgroups

ScalingServerProxyProxyTrainProxyFTColl.ProxyColl.TrainColl.FT
Unscaled66.7655.43 / 50.5055.07 / 45.7659.33 / 54.3165.88 / 60.5062.47 / 60.5766.29 / 62.24
Scaled65.8046.12 / 33.0552.12 / 39.4649.92 / 35.1063.45 / 49.0660.30 / 56.9763.95 / 51.02

Therefore, Theorem 3.1 guarantees the convergence of the unscaled CloakFL configuration, which also maintained server performance and protection effectiveness in the experiments, and we confirmed in Figure 1(b) that the aggregated perturbation remains close to zero even with scaling.

AuthorsOpenReview ↗

Response to Reviewer 59yL (2/5)

W2. Recovery fine-tuning uses only one local epoch at a fixed reduced learning rate, and the evaluation does not test stronger optimization or exploitation of models collected across rounds.

To address this, we evaluated nine conditions combining recovery training epochs (1, 3, 5) and learning rates (1%, 10%, 100% of the base learning rate of 0.1) in the cross-device setting on Tiny-ImageNet (β=0.1). Among these, E=5& LR=0.1 corresponds to the original local training (ProxyTrain), and E=1& LR=0.01 corresponds to the original fine-tuning (ProxyFT).

For FedAvg and FedQSN, we observed several cases where the recovered model outperformed the server model, whereas no such cases were observed for CloakFL. This difference is more pronounced in the Multi Recovery setting.

The Proxy and recovery entries report the maximum / minimum values across the evaluated combinations; a minimum above a server reference means that every evaluated combination exceeded that reference. Values exceeding the Server performance for the corresponding method are highlighted in bold.

Single-Client Recovery

MethodServerProxyE=1& LR=0.001E=1& LR=0.01E=1& LR=0.1E=3& LR=0.001E=3& LR=0.01E=3& LR=0.1E=5& LR=0.001E=5& LR=0.01E=5& LR=0.1
FedAvg35.7035.70 / 35.7036.61 / 35.5536.37 / 32.6513.61 / 2.5636.84 / 35.8032.93 / 27.5815.84 / 5.5336.78 / 35.6830.61 / 25.0814.87 / 3.87
FedQSN29.8727.60 / 23.2031.49 / 29.5831.70 / 26.208.95 / 1.6731.50 / 30.5427.40 / 20.9210.23 / 2.9431.89 / 29.8724.12 / 18.1510.06 / 2.08
CloakFL29.5413.03 / 5.0714.84 / 8.4118.79 / 11.9510.85 / 3.2815.83 / 9.5520.26 / 13.2111.65 / 3.9316.73 / 10.5419.15 / 12.8311.42 / 5.10

Multi-Client Recovery

MethodServerProxyE=1& LR=0.001E=1& LR=0.01E=1& LR=0.1E=3& LR=0.001E=3& LR=0.01E=3& LR=0.1E=5& LR=0.001E=5& LR=0.01E=5& LR=0.1
FedAvg35.7035.70 / 35.7036.79 / 36.2537.59 / 36.6029.46 / 24.7537.24 / 36.6537.58 / 35.7731.44 / 26.2037.33 / 36.7937.37 / 34.9330.56 / 24.60
FedQSN29.8727.11 / 26.1131.21 / 30.6432.96 / 31.8625.44 / 19.6232.22 / 31.6733.69 / 30.5724.07 / 19.0232.53 / 32.1133.00 / 30.1523.27 / 16.84
CloakFL29.5425.15 / 12.6927.12 / 17.2328.21 / 21.8825.99 / 21.5627.70 / 18.8328.55 / 23.9525.14 / 20.4128.21 / 19.9228.64 / 24.0823.75 / 20.17
AuthorsOpenReview ↗

Response to Reviewer 59yL (3/5)

W2. Recovery fine-tuning uses only one local epoch at a fixed reduced learning rate, and the evaluation does not test stronger optimization or exploitation of models collected across rounds.

Across-Round Recovery

To evaluate recovery across multiple rounds, we evaluated across 3R, which selects one proxy from each of three consecutive rounds, and 30% across 3R, which selects 30% of the participating clients in each round, combining nine proxies in total.

Considering the performance improvements from model averaging across rounds discussed in WIMA [1] and IMA [2], we provided the average of the server models from the same three rounds, Server (across 3R), as a reference alongside the recovery results obtained by combining proxies from multiple rounds.

The recovery entries report the maximum / minimum values across the evaluated combinations.

MethodServerServer (across 3R)Coll.Proxy (across 3R)Coll.Proxy (30% across 3R)Coll.Train (across 3R)Coll.Train (30% across 3R)Coll.FT (across 3R)Coll.FT (30% across 3R)
FedAvg35.7039.2139.21 / 39.2139.21 / 39.2129.73 / 14.1639.04 / 32.8539.85 / 38.8440.59 / 39.87
FedQSN29.8734.6432.14 / 30.9231.79 / 31.4023.02 / 3.6732.92 / 26.6535.69 / 34.5336.43 / 35.76
CloakFL29.5431.8725.27 / 13.9529.28 / 25.3523.59 / 17.6631.55 / 26.4028.20 / 21.5032.10 / 28.89

For FedAvg and FedQSN, the maximum recovery performance exceeded Server in several collusion settings. In contrast, for CloakFL, the maximum exceeded Server only in 30% across 3R.

In Coll.FT (30% across 3R), which uses the largest number of proxies, the maximum values for FedAvg and FedQSN exceeded their respective Server (across 3R) values by 1.38 and 1.79 percentage points, and their minimum values also exceeded the corresponding reference values, meaning that every evaluated combination exceeded them. In contrast, the maximum value for CloakFL exceeded the reference value by 0.23 percentage points, but its minimum value was below Server.

[1] Caldarola et al., “Window-Based Model Averaging Improves Generalization in Heterogeneous Federated Learning,” ICCV Workshops, 2023.

[2] Zhou et al., “Understanding and Improving Model Averaging in Federated Learning on Heterogeneous Data,” IEEE TMC, 2024.

Reviewer_ZxAtOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes to use client drift as a convergence-aware perturbation mechanism in proxy model distribution (PMD) to mitigates the risk of model leakage. Theoretically, the paper shows FedAVG-style convergence rates. Empirically, the proposal outperforms recent baselines with interesting ablation studies.

Critical strengths
  • The idea is simple but impactful
  • Good writing
Critical weaknesses
  • Convexity assumption and L-Lipschitz are strong assumptions and usually do not hold in practice. Even though these assumptions are used widely in literature review, the author should provide scenarios or toy experiments to show that the proofs are correct and might be realistic in practice.

  • There might be some gaps between theoretical analysis and the practical implementation. In the analysis, the perturbation is considered as a special error (which can be canceled out) while in the implementation, we design such that such perturbations would not be canceled out by standard averaging (in Multi-Client Recovery experiments). It is unclear to me why it’s guaranteed that this perturbation would not be canceled out.

Questions for authors

See Weaknesses

3 comments
AuthorsOpenReview ↗

Response to Reviewer ZxAt

We sincerely thank the reviewer for the constructive feedback. To address the concerns regarding the identified weaknesses, we provide the following clarifications together with additional experimental results.

W2. In the analysis, the perturbation is considered as a special error (which can be canceled out) while in the implementation, we design such that such perturbations would not be canceled out by standard averaging (in Multi-Client Recovery experiments).

In CloakFL without scaling, the mean of the client drifts assigned in a round is exactly zero, so it satisfies the conditional unbiasedness assumption of Theorem 3.1. This exact cancellation holds when averaging over all clients participating in that round. When averaging only the proxies of a subset of clients, as in the multi-client recovery experiments, complete cancellation of the perturbations cannot be guaranteed.

On Tiny-ImageNet, server performance without scaling was similar to that with scaling, and the performance of both the proxies and the recovered models was below that of the server.

Client-specific scaling was added to improve protection effectiveness in settings such as 20 Newsgroups and to control the trade-off between protection effectiveness and server performance. Exact cancellation is not guaranteed after scaling, but the aggregated perturbation remains close to zero, as shown in Figure 1(b) and the table below.

Tiny-ImageNet

Individual perturbation normAggregated perturbation norm
3.79320.3675

20 Newsgroups

Individual perturbation normAggregated perturbation norm
6.52970.7580

The entries in the table below report the maximum / minimum values across proxy models or recovery sets.

Tiny-ImageNet

ScalingServerProxyProxyTrainProxyFTColl.ProxyColl.TrainColl.FT
Unscaled29.368.51 / 0.5310.22 / 4.7917.58 / 4.6320.08 / 10.7523.88 / 19.0726.03 / 18.79
Scaled29.6910.61 / 4.4612.20 / 1.3518.82 / 11.1323.47 / 15.9523.27 / 18.6427.62 / 22.71

20 Newsgroups

ScalingServerProxyProxyTrainProxyFTColl.ProxyColl.TrainColl.FT
Unscaled66.7655.43 / 50.5055.07 / 45.7659.33 / 54.3165.88 / 60.5062.47 / 60.5766.29 / 62.24
Scaled65.8046.12 / 33.0552.12 / 39.4649.92 / 35.1063.45 / 49.0660.30 / 56.9763.95 / 51.02

Therefore, Theorem 3.1 guarantees the convergence of the unscaled CloakFL configuration, which also maintained server performance and protection effectiveness in the experiments, and we confirmed in Figure 1(b) that the aggregated perturbation remains close to zero even with scaling.

Thank you for your responses. My concern regarding the theoretical assumptions is not sufficiently addressed. it is unclear whether these assumptions hold in realistic settings, raising questions about the practical applicability and generalizability of the theoretical results.

AuthorsOpenReview ↗

Response to Reviewer ZxAt

We thank the reviewer for clarifying the concern. While we acknowledge that convexity and smoothness assumptions may not hold for general deep neural networks, these are widely adopted in convergence analyses of federated learning, including FedAvg. Our theoretical results provide rigorous guarantees under these standard assumptions. Furthermore, our extensive experiments on non-convex CNN and Transformer architectures demonstrate that CloakFL remains effective beyond the theoretical setting, supporting its practical applicability. We believe that our theoretical analysis and empirical evaluations together provide complementary evidence for the effectiveness of CloakFL.

Reviewer_9ojKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper has proposed CloakFL, a method that protects the global model in FL from being leaked or stolen by the participating clients.

In the traditional federated learning, the server shares the global model with the clients and client can launch the local training. This can create a security risk because a malicious client could copy and possibly redistribute the model.

To address this challenge, the authors have proposed sharing a degraded versions of the central server model, called proxy models, instead of the original model. The main challenge is that degrading these models can also hurt the training process and reduce the effectiveness or the performance of the final model.

CloakFL proposes to solve this problem by using information from the previous client updates to generate perturbations. These perturbations may reduce the performance of individual proxy models but will have a smaller impact on the global model after weight aggregation.

The authors have evaluated CloakFL under different federated learning settings. They compare their approach with FedAvg and FedQSN . They have also evaluated whether malicious clients can recover the original model through additional training or collaboration.

Critical strengths

1- I think the main strength of this paper is the idea of using client drift to generate perturbations instead of relying only on random noise, masking, or quantization. This is an interesting direction because it uses information already available during federated training rather than introducing a completely separate protection mechanism.

2- Another strength is that the authors go beyond measuring the accuracy of degraded proxy models. As shown in table 2 and 3, this study considers whether individual clients can recover model performance through additional training or not, and if multiple clients can improve the overall performance by combining their respective models.

Critical weaknesses
  1. Limited evaluation against recovery attacks

My main concern is that the paper focuses on protecting the global model from malicious clients but the recovery experiments do not fully test the robustness of the defense.

For single-client attacks the authors mainly consider local training and one extra epoch of fine-tuning with a lower learning rate.. In reality an attacker would not be limited to these approaches. They could train for more epochs test different learning rates or gather proxy models across multiple communication rounds to reconstruct the original model.

The paper also explores attacks with clients, which is helpful.. These experiments use fairly simple strategies like averaging and fine-tuning involving at most 30% of the clients working together.

These experiments are a beginning but they don't go far enough to show how well CloakFL stands up against more determined attackers. Since model protection is the goal of the paper I think stronger recovery attacks are needed before we can confidently assess the security of CloakFL.

  1. Gap between the analysis and the actual experiments

The convergence analysis, in the paper is useful but I’m not sure how much it reflects what actually happens in practice.

For example the analysis assumes that all clients participate in every round and that the loss function is strongly convex.. The experiments use CNNs and Transformers—models that are not strongly convex—and only a fraction of clients take part in each round.

More important the convergence guarantee depends on the aggregated perturbations being unbiased. However CloakFL scales the perturbations differently per client. The authors admit this does not ensure perfect cancellation.

So it’s not clear that the perturbations generated by CloakFL meet the assumptions used in the theory.

The authors should clarify this connection. Without an explanation it's hard to know whether the convergence guarantee actually applies to the proposed method or only holds under idealized conditions.

Questions for authors
  1. Stronger recovery attacks

The recovery experiments mostly use one epoch of fine-tuning with a small learning rate.. What if an attacker runs training for more epochs? or Tries different learning rates and optimizers? Do we have ablations for that?

Since the goal of CloakFL is to protect the global model it’s important to know how much of the model’s performance an attacker might actually recover if they put in more effort. More training and more computing power could make a difference. I think this should be tested.

  1. Recovery across multiple communication rounds

The paper looks at attacks where several clients team up in one communication round.. An attacker might also save the proxy models they get over many rounds and try to combine them later.

Have the authors thought about this? Could combining models from rounds help an attacker get a better guess of the global model? This seems important especially because CloakFL uses past information to create its perturbations. It might make the attack stronger over time.

  1. Connection between the theory and the actual algorithm

The convergence theorem says the aggregated perturbations must be conditionally unbiased.. In Section 4.2 the paper says scaling perturbations differently for each client doesn’t mean the perturbations will cancel out exactly.

So do the perturbations created by CloakFL actually meet the conditions of Theorem 3.1? If not then what does the convergence guarantee mean for this method? I’d like to understand how well the theory matches what actually happens in practice.

  1. Impact of the number of participating clients

The experiments use 20 or 100 clients with 10 joining each round.

What if fewer or more clients joined each round? Since CloakFL depends on perturbations canceling out during aggregation I’d expect the number of clients to affect both how well the model is protected and how fast it converges.

It would be useful to know how sensitive the method is, to this. Does it work well with 5 clients?. Break down with 50? That kind of info would help people decide how to set it up.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 9ojK (1/5)

We sincerely thank the reviewer for the constructive feedback. To address the concerns regarding the identified weaknesses, we provide the following clarifications together with additional experimental results.

W1. For single-client attacks the authors mainly consider local training and one extra epoch of fine-tuning with a lower learning rate..

To address this, we evaluated nine conditions combining recovery training epochs (1, 3, 5) and learning rates (1%, 10%, 100% of the base learning rate of 0.1) in the cross-device setting on Tiny-ImageNet (β=0.1). Among these, E=5& LR=0.1 corresponds to the original local training (ProxyTrain), and E=1& LR=0.01 corresponds to the original fine-tuning (ProxyFT).

For FedAvg and FedQSN, we observed several cases where the recovered model outperformed the server model, whereas no such cases were observed for CloakFL. This difference is more pronounced in the Multi Recovery setting.

The Proxy and recovery entries report the maximum / minimum values across the evaluated combinations; a minimum above a server reference means that every evaluated combination exceeded that reference. Values exceeding the Server performance for the corresponding method are highlighted in bold.

Single-Client Recovery

MethodServerProxyE=1& LR=0.001E=1& LR=0.01E=1& LR=0.1E=3& LR=0.001E=3& LR=0.01E=3& LR=0.1E=5& LR=0.001E=5& LR=0.01E=5& LR=0.1
FedAvg35.7035.70 / 35.7036.61 / 35.5536.37 / 32.6513.61 / 2.5636.84 / 35.8032.93 / 27.5815.84 / 5.5336.78 / 35.6830.61 / 25.0814.87 / 3.87
FedQSN29.8727.60 / 23.2031.49 / 29.5831.70 / 26.208.95 / 1.6731.50 / 30.5427.40 / 20.9210.23 / 2.9431.89 / 29.8724.12 / 18.1510.06 / 2.08
CloakFL29.5413.03 / 5.0714.84 / 8.4118.79 / 11.9510.85 / 3.2815.83 / 9.5520.26 / 13.2111.65 / 3.9316.73 / 10.5419.15 / 12.8311.42 / 5.10

Multi-Client Recovery

MethodServerProxyE=1& LR=0.001E=1& LR=0.01E=1& LR=0.1E=3& LR=0.001E=3& LR=0.01E=3& LR=0.1E=5& LR=0.001E=5& LR=0.01E=5& LR=0.1
FedAvg35.7035.70 / 35.7036.79 / 36.2537.59 / 36.6029.46 / 24.7537.24 / 36.6537.58 / 35.7731.44 / 26.2037.33 / 36.7937.37 / 34.9330.56 / 24.60
FedQSN29.8727.11 / 26.1131.21 / 30.6432.96 / 31.8625.44 / 19.6232.22 / 31.6733.69 / 30.5724.07 / 19.0232.53 / 32.1133.00 / 30.1523.27 / 16.84
CloakFL29.5425.15 / 12.6927.12 / 17.2328.21 / 21.8825.99 / 21.5627.70 / 18.8328.55 / 23.9525.14 / 20.4128.21 / 19.9228.64 / 24.0823.75 / 20.17
AuthorsOpenReview ↗

Response to Reviewer 9ojK (2/5)

Q2. Could combining models from rounds help an attacker get a better guess of the global model?

Across-Round Recovery

To evaluate recovery across multiple rounds, we evaluated across 3R, which selects one proxy from each of three consecutive rounds, and 30% across 3R, which selects 30% of the participating clients in each round, combining nine proxies in total.

Considering the performance improvements from model averaging across rounds discussed in WIMA [1] and IMA [2], we provided the average of the server models from the same three rounds, Server (across 3R), as a reference alongside the recovery results obtained by combining proxies from multiple rounds.

The recovery entries report the maximum / minimum values across the evaluated combinations.

MethodServerServer (across 3R)Coll.Proxy (across 3R)Coll.Proxy (30% across 3R)Coll.Train (across 3R)Coll.Train (30% across 3R)Coll.FT (across 3R)Coll.FT (30% across 3R)
FedAvg35.7039.2139.21 / 39.2139.21 / 39.2129.73 / 14.1639.04 / 32.8539.85 / 38.8440.59 / 39.87
FedQSN29.8734.6432.14 / 30.9231.79 / 31.4023.02 / 3.6732.92 / 26.6535.69 / 34.5336.43 / 35.76
CloakFL29.5431.8725.27 / 13.9529.28 / 25.3523.59 / 17.6631.55 / 26.4028.20 / 21.5032.10 / 28.89

For FedAvg and FedQSN, the maximum recovery performance exceeded Server in several collusion settings. In contrast, for CloakFL, the maximum exceeded Server only in 30% across 3R.

In Coll.FT (30% across 3R), which uses the largest number of proxies, the maximum values for FedAvg and FedQSN exceeded their respective Server (across 3R) values by 1.38 and 1.79 percentage points, and their minimum values also exceeded the corresponding reference values, meaning that every evaluated combination exceeded them. In contrast, the maximum value for CloakFL exceeded the reference value by 0.23 percentage points, but its minimum value was below Server.

[1] Caldarola et al., “Window-Based Model Averaging Improves Generalization in Heterogeneous Federated Learning,” ICCV Workshops, 2023.

[2] Zhou et al., “Understanding and Improving Model Averaging in Federated Learning on Heterogeneous Data,” IEEE TMC, 2024.

AuthorsOpenReview ↗

Response to Reviewer 9ojK (3/5)

Q3. So do the perturbations created by CloakFL actually meet the conditions of Theorem 3.1?

In CloakFL without scaling, the mean of the client drifts assigned in each round is exactly zero, so the aggregated perturbation also cancels out completely. Therefore, this configuration satisfies the conditional unbiasedness assumption of Theorem 3.1. On Tiny-ImageNet, server performance without scaling was similar to that with scaling, and the performance of both the proxies and the recovered models was below that of the server.

Client-specific scaling was added to improve protection effectiveness in settings such as 20 Newsgroups and to control the trade-off between protection effectiveness and server performance. Exact cancellation is not guaranteed after scaling, but the aggregated perturbation remains close to zero, as shown in Figure 1(b) and the table below.

Tiny-ImageNet

Individual perturbation normAggregated perturbation norm
3.79320.3675

20 Newsgroups

Individual perturbation normAggregated perturbation norm
6.52970.7580

The entries in the table below report the maximum / minimum values across proxy models or recovery sets.

Tiny-ImageNet

ScalingServerProxyProxyTrainProxyFTColl.ProxyColl.TrainColl.FT
Unscaled29.368.51 / 0.5310.22 / 4.7917.58 / 4.6320.08 / 10.7523.88 / 19.0726.03 / 18.79
Scaled29.6910.61 / 4.4612.20 / 1.3518.82 / 11.1323.47 / 15.9523.27 / 18.6427.62 / 22.71

20 Newsgroups

ScalingServerProxyProxyTrainProxyFTColl.ProxyColl.TrainColl.FT
Unscaled66.7655.43 / 50.5055.07 / 45.7659.33 / 54.3165.88 / 60.5062.47 / 60.5766.29 / 62.24
Scaled65.8046.12 / 33.0552.12 / 39.4649.92 / 35.1063.45 / 49.0660.30 / 56.9763.95 / 51.02

Therefore, Theorem 3.1 guarantees the convergence of the unscaled CloakFL configuration, which also maintained server performance and protection effectiveness in the experiments, and we confirmed in Figure 1(b) that the aggregated perturbation remains close to zero even with scaling.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Follow-up

Dear Reviewers,

We would like to kindly follow up on our rebuttal and highlight two key concerns shared across the reviews. We have provided detailed clarifications and additional experimental results addressing these concerns, which we briefly summarize below.

1. Gap between Assumption 3.3 and practical implementation.

As clarified in our rebuttal, the unscaled version of CloakFL satisfies the conditional unbiasedness requirement in Assumption 3.3, as the aggregated perturbations cancel out. While client-specific scaling does not guarantee exact cancellation, our empirical results show that the aggregated perturbation remains small relative to individual perturbations (0.37 vs. 3.79 on Tiny-ImageNet and 0.76 vs. 6.53 on 20 Newsgroups). Furthermore, our additional ablation studies demonstrate that even the unscaled version maintains effective model protection while preserving server performance, supporting the practical relevance of our theoretical analysis.

2. Robustness against diverse recovery attack scenarios.

We substantially extended our evaluation to include nine combinations of recovery epochs and learning rates, cross-round recovery attacks, varying numbers of participating clients, and exhaustive evaluations of individual clients and colluding coalitions.

Although the maximum recovery performance occasionally exceeds the server performance in certain collusion scenarios, such successful recovery occurs in only a very small fraction of colluding sets. In our exhaustive Tiny-ImageNet evaluation, only 1.67% of all possible colluding sets achieved server-level performance under collusion fine-tuning, compared with 100% for FedQSN, the existing PMD baseline. This highlights that, even when recovery is possible for a few favorable coalitions, CloakFL remains resistant to recovery across the vast majority of evaluated combinations.

We hope these additional clarifications and experiments help address the main concerns regarding the theoretical assumptions and practical robustness of CloakFL.

We would greatly appreciate it if you could revisit our rebuttal and share any further feedback or remaining concerns.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.