FuseFL: One-Shot Federated Learning through the Lens of Causality with Progressive Model Fusion

One-shot Federated Learning (OFL) significantly reduces communication costs in FL by aggregating trained models only once. However, the performance of advanced OFL methods is far behind the normal FL. In this work, we provide a causal view to find that this performance drop of OFL methods comes from the isolation problem, which means that local isolatedly trained models in OFL may easily fit to spurious correlations due to the data heterogeneity. From the causal perspective, we observe that the spurious fitting can be alleviated by augmenting intermediate features from other clients. Built upon our observation, we propose a novel learning approach to endow OFL with superb performance and low communication and storage costs, termed as FuseFL. Specifically, FuseFL decomposes neural networks into several blocks, and progressively trains and fuses each block following a bottom-up manner for feature augmentation, introducing no additional communication costs. Comprehensive experiments demonstrate that FuseFL outperforms existing OFL and ensemble FL by a significant margin. We conduct comprehensive experiments to show that FuseFL supports high scalability of clients, heterogeneous model training, and low memory costs. Our work is the first attempt using causality to analyze and alleviate data heterogeneity of OFL.

Paper

Similar papers

Peer review

Reviewer twUP6/10 · confidence 4/52024-06-12

Summary

This paper proposes FuseFL to address the non-IID data in one-shot federated learning (OFL). Specifically, FuseFL is inspired by the lens of casuality, and fuses each layer of the global model step by step to learn invariant features. Extensive experiments validate that FuseFL achieves SOTA accuracy under various non-IID settings.

Strengths

S1: The proposed algorithm is novel and well-motivated. S2: Extensive experiments validate the effectiveness of FuseFL.

Weaknesses

W1: Since FuseFL still communicates multiple rounds, I think it belongs to few-shot FL instead of one-shot FL. One-shot FL should have only one communication round. Besides, FedAvg (or other FL algorithms) with the same number of communication rounds should be compared as baselines to ensure a fair comparison. W2: More experimental setups can be done to further support the effectiveness of the proposed algorithm. For example, in a previous FL benchmark on non-IID data [1], the quantity-based label skew settings are challenging, which are not tested in the current version. Also, in the section of scalability, the authors only test up to 50 clients. I suggest the authors to test on even more clients, e.g. 500 clients, to validate the scalability. [1] Li, Qinbin, et al. "Federated learning on non-iid data silos: An experimental study." 2022 IEEE 38th international conference on data engineering (ICDE). IEEE, 2022.

Questions

Please see weaknesses. I think this paper is sound and interesting. I'd like to see more experiments to further verify the effectiveness of the proposed algorithm.

Rating

6

Confidence

4

Soundness

2

Presentation

3

Contribution

3

Limitations

Please see weaknesses.

Reviewer V6z47/10 · confidence 5/52024-07-11

Summary

The authors identify the isolation problem as being the root cause of low accuracy in OFL. The isolation problem arises as the clients locally overfit to spurious features observed in their own local datasets in the absence of knowledge from other clients. To better learn invariant features, the authors propose progressive model fusion called FuseFL which fuses client local models block by block. Crucially, this can be done without increasing the total communication cost, however at the cost of additional rounds.

Strengths

- The authors address an important problem in OFL research in a novel way - The problem and its solution is clearly presented - The experimental evaluation is diverse, in particular, the authors also show the memory cost of fusion and experiments with heterogeneous models

Weaknesses

- An important baseline [1] is missing from the evaluations which has been shown to improve over the uniform ensemble considered as upper bound in this paper. The CoBoosting algorithm in [1] generates the weights of the ensemble simultaneously with synthetic data to transfer the knowledge to a single global model and is the recent SOTA. Notably, it is a data-free ensemble distillation method in contrast to Table 1. - No evaluations are shown with higher heterogeneity levels, such as a = 0.01 or a = 0.05 which is a common setting in OFL papers [1,2,3]. This prevents a comprehensive comparison with prior work. - The authors have not discussed important limitations of their work. In particular, the security aspect under multiple rounds of communication. OFL algorithms are recognised not just for their low communication costs but also their lower vulnerability to attacks, thanks to the single communication round. By introducing multiple rounds of communication, FuseFL increases vulnerability to attacks. - Similarly, FuseFL appears to increase client side overheads by letting clients run training K times instead of once as in standard OFL. While the computation cost appears to decrease with K, it is nevertheless higher than client side overheads in standard OFL approaches which mostly add server side overheads. Hence, a comprehensive discussion of these limitations is expected to provide a complete picture of tradeoffs. [1] R. Dai, Y. Zhang, A. Li, T. Liu, X. Yang, and B. Han. Enhancing one-shot federated learning through data and ensemble co-boosting. In The Twelfth International Conference on Learning Representations, 2024. [2] C. E. Heinbaugh, E. Luz-Ricca, and H. Shao. Data-free one-shot federated learning under very high statistical heterogeneity. In The Eleventh International Conference on Learning Representations, 2023. [3] Jhunjhunwala, D., Wang, S. and Joshi, G., 2024, April. FedFisher: Leveraging Fisher Information for One-Shot Federated Learning. In the International Conference on Artificial Intelligence and Statistics (pp. 1612-1620). PMLR.

Questions

The reviewer is positive about this work and is willing to increase the score provided the authors evaluate against Co-Boosting and include a higher heterogeneity level in their overall assessment. While a thorough empirical assessment of the security of the proposed approach is not deemed necessary, the paper must at least discuss the security aspects along with the client side overheads to better inform the readers of the inherent tradeoffs in FuseFL.

Rating

7

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

No additional limitations than those discussed in the sections above. The authors have well addressed other limitations of their approach involving the memory footprint of the fused model, applicability to model heterogeneity, etc. through their empirical evaluations.

Reviewer FVeg5/10 · confidence 2/52024-07-19

Summary

The authors provide a causality viewpoint for the data heterogeneity problem in the training of one-shot fed-avg. A block fusing mechanism in the training to provide more information aggregation is designed and the results show a significant improvement for the common tests in ResNet and Cifar.

Strengths

Pros: 1. the one-shot FL is an interesting and important problem. The trials on the non-IID processing is difficult but significant, which can be further expanded to more industrial scenes. 2. The idea is well-represented and the performances are great according to the providing verification.

Weaknesses

Cons: 1. the cost of the multiple-block communication needs to be considered. In the computation, there might be a huge amount of blocks, we need to keep connections or wait for the block communications in the practical applications. 2. the explanation part needs to be more rigorous. Please use theorems or propositions for clear statements. I believe the current version can not provide a convincing theoretical explanation from the casual viewpoint. 3. More large-scale dataset and models need to be considered. The introduction brings a large scope for LLM FL training but not experiments are designed for it.

Questions

1. what is the definition of "block"/"features"? I think it will influence the results from the theoretical point of view. 2. how to deal with the scenes where the disconnection occurs in training, can the system still work (like dynamic topologies in the block communications) 3. please show the error bar of the experiments since it will not perform significantly better than Ensemble FL for a few of the cases. 4. I'm also interested in how the model works on the Transformers-based framework, and which part we should aggregate for these models. I also wonder if the method can work on the prompt for LLMs. 5. Please verify the non-IID data-generating processes and whether and how they align with the explanation part.

Rating

5

Confidence

2

Soundness

2

Presentation

3

Contribution

2

Limitations

N/A.

Reviewer mHC67/10 · confidence 4/52024-07-20

Summary

This paper proposes a method FuseFL that aims to perform one-shot FL by progressively fusing models from multiple clients. Their motivation is that within each local client, a model might learn spurious features due to the underlying spurious correlations, adversarial attacks, and shortcuts. However, at the global level, the fusion process would eliminate those spurious features and emphasize on the core/invariant features. They propose a strategy of one-shot fusion of models from multiple clients, followed by experimental results. They also include arguments using mutual information over a Markov chain that use the Data-Processing Inequality. Experimental results show benefits over other methods in terms of accuracy of the model. UPDATE: Increased soundness to 3, Rating to 7 Accept

Strengths

The paper addresses a very interesting problem at the intersection of federated learning and spurious correlations. It is an interesting observation that spurious correlations prevalent at the local model would go away at the global stage by appropriate fusion. The idea of one-shot fusion is interesting. They have made an attempt to explain what is going on using information theory and information bottleneck principles. Experiments compare their strategy with other strategies across several datasets.

Weaknesses

1. I have some concerns on the clarity of the theoretical analysis here, which if resolved, will increase my rating because I otherwise liked the idea of the paper. I believe Nuisance is mathematically different from Spurious Features (you are assuming them to be the same here)? Nuisance [2] is N which has no information about the true label Y. But, typically in spurious correlation papers, Rspu is something that has information about Y even though they are not causally related, e.g., in Waterbirds dataset, many waterbirds are photographed in front of water background. So, water background (B) naturally has a correlation with bird label (Y) and is not nuisance. But, it is not desirable to use B in the model since it performs poorly on minority groups. Here B is Rspu but not Nuisance. The authors also write I(Rspu;H)=0 means spurious feature is not being used. This would also hold for nuisance but not spurious feature as in the example above. One wants the model to not mechanically use the background B. But, what it actually uses instead of the background can still have correlations with the model outputs, and hence mutual information can still be there. I think what you call Rspu is actually Nuisance at a global level. You actually assume that for the global dataset, Rspu satisfies the definition of nuisance, i.e., being independent with Y, but is dependent at a local level? Could you include a clear notation and definition section where you clarify your notations and independence assumptions on Rspu and other terms at global and local levels? 2. Experiments: In spurious correlation papers, one is often interested in worst-group/minority group accuracy, i.e., waterbirds on land background. Overall accuracy even decreases sometimes. Could you comment on this in the context of your experiments? How are the local accuracies? 3. There is an overreliance on causality literature including citing Pearl. However, I feel the strategy and the novelty are not that reliant on causality. 4. The fusion strategy is not properly described. Everything points to one equation (5) which is not described clearly.

Questions

See Weaknesses above. Could you include a clear notation and definition section where you clarify your notations and independence assumptions on Rspu and other terms at global and local levels? If I(Hk; Rspu)=0 is necessary/sufficient? This is not really a causality paper. I believe the overreliance on causality is not necessary here. The fusion strategy is not properly described. Everything points to one equation (5) which is not described clearly.

Rating

7

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

Discussed above

Authorsrebuttal2024-08-14

Thanks for your helpful comments and new related works

Dear Reviewer #mHC6, Thanks for providing these related works [1,2]. Removing local biased feature is an interesting idea to enhance the global fairness. We will add these discussions and related works into our revision. Also, we will modify the paper to highlight the information bottleneck/mutual information and the title. Thanks a lot for your suggestions. Best regards and thanks, > ***Reference*** > > [1] Yahya H Ezzeldin, Shen Yan, Chaoyang He, Emilio Ferrara, and A Salman Avestimehr. Fairfed: Enabling group fairness in federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pp. 7494–7502, 2023. > > [2] F. Hamman and S. Dutta, "Demystifying Local & Global Fairness Trade-offs in Federated Learning Using Partial Information Decomposition,” International Conference on Learning Representations (ICLR 2024).

Reviewer twUP2024-08-08

Rebuttal ack

I appreciate the authors' response. Considering other reviews, I raise the score to "weak accept".

Authorsrebuttal2024-08-08

Thanks for your swift reply and raising the score

Dear Reviewer #twUP, We're grateful for your quick feedback during this busy period. We deeply appreciate your consideration in raising the score. We remain open and ready to delve into any more questions or suggestions you might have. Your constructive comments have significantly contributed to the refinement of our work. Best regards and thanks

Reviewer V6z42024-08-10

The reviewer thanks the authors for their responses. The authors have well addressed the concerns regarding higher heterogeneity level, missing baseline and computational cost of FuseFL. The reviewer will hence increase the score. However, as was mentioned in the review, the reviewer highly recommends adding a discussion section describing the security vulnerabilities and possible mitigations. This should not be seen as a weakness but something that will enhance the quality of the paper.

Authorsrebuttal2024-08-11

Thanks for your swift reply and raising the score

Dear Reviewer #V6z4, Thank you for your prompt response during this busy period. We deeply appreciate your consideration in raising the score. According to the **security issues**, we provide following **possible mitigations** and add into our revision: - **Adversarial attacks:** Some malicious clients might upload adversarial modules or backdoored modules that are used to misguide the aggregated model to generate incorrect or handcrafted predictions. For these attacks, the possible solution is to **detect and reject such malicious uploading** also through the lens of causality. Specifically, some images with the invariant features can be fed into the uploaded modules to see whether the output feature can be used to correctly classify images; - **Model inversion or Membership attack:** Some malicious clients or the server may consider to conduct model inversion or membership attack to obtain the raw data of clients, thus threatening the user privacy. In this case, the learned module can be protected with differential privacy to enhance its security. We remain open and ready to delve into any more questions or suggestions you might have. Your constructive comments have significantly contributed to the refinement of our work. Best regards and thanks,

Reviewer mHC62024-08-12

Thanks for your response

I have read the response. Based on the proposed edits to the theoretical section, I increased my soundness score to 3, and also update my rating to Accept 7.

Reviewer mHC62024-08-12

Suggestion on Title

Only one more thing: I feel that "through the Lens of Causality" is not well suited for the title of the paper since the authors also acknowledge that the over-reliance on causality is not needed and this is not a causality paper. Something else to highlight the information-bottleneck-style analysis will be better.

Authorsrebuttal2024-08-12

Thanks for your swift reply and raising the score

Dear Reviewer #mHC6, We're grateful for your quick feedback. We deeply appreciate your consideration in raising the score. Thank you for your suggestion. We will try to give a reasonable title, from the perspective of the information bottleneck and the invariant/spurious features. We remain open and ready to delve into any more questions or suggestions you might have. Your constructive comments has significantly contributed to the refinement of our work. Best regards and thanks,

Reviewer mHC62024-08-13

Continued

In the paper: "causal" could be replaced with sequential or temporal In the title: The role of information bottleneck/mutual information could be highlighted Related works: Once presented as a local-global tussle, other relevant works within federated learning which also look at this tussle should be acknowledged. For instance, the problem of local-global fairness seems to be quite aligned with this work, and a fusion approach is indeed quite valuable. Some suggested references: [1] Yahya H Ezzeldin, Shen Yan, Chaoyang He, Emilio Ferrara, and A Salman Avestimehr. Fairfed: Enabling group fairness in federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pp. 7494–7502, 2023. [2] F. Hamman and S. Dutta, "Demystifying Local & Global Fairness Trade-offs in Federated Learning Using Partial Information Decomposition,” International Conference on Learning Representations (ICLR 2024). Good work!

Reviewer FVeg2024-08-13

Thank you for your comprehensive explanations and diligent efforts during the rebuttal stage. Your responses have addressed some of my concerns.

Authorsrebuttal2024-08-13

Thanks for your reply and insightful comments

Dear Reviewer #FVeg, We're grateful for your feedbacks during this busy period. We will remain open and ready to delve into any more questions or suggestions you might have until the last moment. Your constructive comments and questions has significantly enhanced the quality of our work. Best regards and thanks

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC