Towards Personalized Federated Learning via Heterogeneous Model Reassembly

This paper focuses on addressing the practical yet challenging problem of model heterogeneity in federated learning, where clients possess models with different network structures. To track this problem, we propose a novel framework called pFedHR, which leverages heterogeneous model reassembly to achieve personalized federated learning. In particular, we approach the problem of heterogeneous model personalization as a model-matching optimization task on the server side. Moreover, pFedHR automatically and dynamically generates informative and diverse personalized candidates with minimal human intervention. Furthermore, our proposed heterogeneous model reassembly technique mitigates the adverse impact introduced by using public data with different distributions from the client data to a certain extent. Experimental results demonstrate that pFedHR outperforms baselines on three datasets under both IID and Non-IID settings. Additionally, pFedHR effectively reduces the adverse impact of using different public data and dynamically generates diverse personalized models in an automated manner.

Paper

Similar papers

Peer review

Reviewer xdZD5/10 · confidence 4/52023-06-23

Summary

The paper proposes to use the recently published model reassembly technique (NeurIPS 2022) to obtain personalized models through federated learning. At each round, the server collects the current models from the clients and uses reassembly to generate new candidate models, potentially training some stitching layers using a public datasets. Then the closest candidate model (again similarity is evaluated through the public dataset) is sent back to the client who can (later) train its own model distilling knowledge from the candidate model. By using the public dataset only for a few selected operations, the proposed scheme should be more robust to deviations from the training dataset and the public one.

Strengths

* The idea to use reassembly for federated learning is a novel one to the best of my knowledge * Experimental results are promising and the robustness to the choice of the public dataset is definitely an important plus of the proposed approach.

Weaknesses

* In the proposed scheme, the server does not keep historical aggregate information about the training, as for example it does under FedAvg by storing the last version of the shared model. Historical information is rather kept at the client, which at each round performs a local training with knowledge distillation from the candidate model selected by the server during the previous communication round to which the client participated. As a consequence the method does not seem suited for large-scale cross-device settings where clients may be selected only a few times. The authors have considered clients' sampling rates between 1/3 and 1/10. I expect performance to decrease significantly for lower rates. * The proposed solution requires the server to maintain the identity of the each client (to be able to send back the relevant candidate model). This prevents the applicability of privacy-preserving techniques like secure aggregation. Note that there are other personalized approaches which do not have this constraints (e.g., Ditto, FedEM,...) * Computational overhead. If I understood correctly, the server needs to train the stitching part for every possible client/candidate-model pair, i.e., to train in total BM models, which poses a significant load on the server. * Complexity of the proposed solution. It would have been good to perform an ablation study to evaluate if all pFedHR steps are really needed. For example, what if the clients' models are directly compared and the closest one is sent as candidate model to the client without performing any reassembly and stitching? * The candidate model can be more complex than the client's model. There is then an implicit assumption that, while the client has selected a given model size for example on the basis of its computational and memory capabilities, it is still able to use a more complex model for knowledge distillation at training time * The comparison with the previous literature is not always clear. Two examples: 1. a limitation of previous literature would be that "the averaging process significantly diminishes the characteristics of individual local models" I found this sentence too vague. 2. "however, FedDF trains a global model with different settings compared with our approach." Again, this is too vague, what is the difference with FedDF in a few words? * compute * the authors checked the compute checkbox but I was not able to find any information about computation in the paper or in the supplementary material * reproducibility * while the code is provided there is no readme file about how to use it and how to reproduce the results in the paper. * minors: * footnote numbers should go after punctuation marks * report the number of clients for table 4 * typos: bettwen and cadidates

Questions

* in the homogeneous setting, candidate model generation reduces to simply swapping layers across clients' models? or is it still possible to obtain different models' architectures (e.g. with a higher number of layers than any of the original models?) * are results reported averages over different experiments or single-run experiments? I am also asking because some of the reported differences are really small (e.g., in table 2 pFedHR outperforms sota by less than 1 percentage point in many configurations)

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

I think the paper should have discussed the following limitations (see corresponding weaknesses above) * performance under low clients' sampling rate * the server keeps track of clients' update * Computational overhead * The clients need to work with models more complex than its own.

Reviewer nvYR6/10 · confidence 4/52023-07-04

Summary

The authors designed a method to train a model with FL when clients have heterogeneous model architectures. They designed a model reassembly technique that stitches together parts of DNNs. pFedHR also creates personalised models for each client without requiring server-side data or explicit human guidance.

Strengths

- Heterogeneous FL is an important problem - The paper is well-written and easy to understand - The authors compared their approach with number of existing algorithms

Weaknesses

- The reason to stitch together heterogeneous architectures is not very well motivated. There are many other approaches that aim to learn personalised model for a set of clients with heterogeneous capabilities. For example FjORD [https://arxiv.org/abs/2102.13451] train a number of subnets using adaptive dropout, HeteroFL [https://arxiv.org/abs/2010.01264] is another model where a superset is trained and submodes are used to address heterogeneity in FL, FedRolex [https://arxiv.org/abs/2212.01548], or [https://arxiv.org/abs/2210.16105] all use weight sharing to address similar tasks. The authors should discuss what is the main advantage/motivation of having heterogeneous architectures stitched together over other approaches where the computational complexity of a model can be scaled up/down, and possibly compare with some of the above methods. - Most experiments were done with few clients (N=12 up to 100). This is very small and might be unrealistic in real-=world applications where we might have million of clients participating in FL. It would be great to show how things scale up.

Questions

- Maybe motivate better why should we personalise the whole architecture and not just the model weights ? - In most cases, it would make sense to generate a few models for different device Tiers (e.g., low-tier old devices, mid-tier, high-tier). It is unclear why automatically generating client-specific personalised models is a good idea (we will have to ensure and measure the quality thousand of different architectures in production).

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

N/A

Reviewer nXSK8/10 · confidence 5/52023-07-07

Summary

In this paper, the authors introduce a technique that tackles the challenge of enabling collaboration among client models with different network structures in federated learning. Unlike traditional knowledge distillation (KD)-based approaches, the proposed model involves dividing the heterogeneous models into distinct parts and subsequently reassembling them. This unique reassembling approach fosters model diversity, facilitating personalized local training at each client. The proposed technique alleviates the adverse impact of utilizing public datasets on the server. By providing experimental results, the authors establish the effectiveness of their proposed model.

Strengths

- Overall, the paper is well-written and easy to follow. The motivations behind the research are clearly articulated, and the experimental results presented are both sufficient and convincing. - The introduction of the novel reassembling technique to federated learning is a significant contribution. This approach opens up new avenues for achieving model personalization, which has the potential to inspire further exploration by researchers in the field. - The authors' inclusion of an empirical analysis that explores the negative impact of utilizing heterogeneous public data on the server is a notable and valuable aspect of this work. Moreover, the proposed reassembling solution effectively mitigates this issue, providing a practical and effective resolution.

Weaknesses

- In the experiments, the authors use four self-designed CNN-based models for saving computational resources. However, there are some lightweight models, such as MobileNets, used in previous FL work. Incorporating these lightweight models would enhance the comprehensiveness of the experimental evaluation and provide a broader perspective on the proposed approach. - Figure 1 is a little hard to read, and I suggest the authors use one setting as an example to demonstrate the challenge of using heterogeneous public data and put other results in Section 4.4. - Adding a readme file to the source code package would be a useful addition.

Questions

no

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authers have provided some discussions.

Reviewer nMvK7/10 · confidence 4/52023-07-07

Summary

The authors design a novel approach to make it possible for clients equipped with different model structures to cooperate in the federated learning framework. Specifically, the server will reassemble models into different parts and assemble them together. After that, they propose a similarity-based approach to match the most fitted model with the clients’ models and distribute them back to the clients respectively. To emphasize the advantages over the KD-based approaches, the authors discuss different perspectives of using public datasets. Finally, they address the effectiveness of the algorithm using experiment results under the IID and non-IID setting compared with other baselines.

Strengths

1. This paper presents a new way to achieve personalized federated learning using heterogeneous model reassembly, which is significantly different from existing work. 2. The authors aim to design a new model to alleviate the issue of performance drop caused by introducing public data on the server, especially when its data distribution is different from that of clients. 3. The authors conduct extensive experiments on different settings, including 12 clients, 100 clients, IID, non-IID, and public data with labels and without labels. The experimental results demonstrate the effectiveness of the proposed model.

Weaknesses

1. The training of stitching layers is not significantly clear. Are they trained with the other parts of the networks or trained separately? The authors can provide a description of how the networks are trained after they are stitched together. 2. Since the generated candidates may change for different runs, are the results averaged by multiple runs? 3. The font size used in Figure 1 is too small.

Questions

Please see the weaknesses above.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

4 excellent

Limitations

The authors discussed the limitations of the proposed work.

Reviewer nvYR2023-08-10

Thanks for the detailed response. > The models, FjORD, HeteroFL, FedRolex, and AsyncDrop, all have a strong constraint, i.e., the client models are required to be submodels of the global model. This assumption might be too strong in real-world scenarios. Can you elaborate on that ? why is this too strong ? What are the practical limitations that would make the above methods not feasible in the real world ? > Client number vs accuracy It seems the accuracy with higher number of clients is dropping. While each client has less data, the FL algorithm still goes through all the data in the dataset. So would you say this approach is scalable when we have million of possible clients in the real-world ?

Authorsrebuttal2023-08-11

Thanks for your prompt and insightful feedback.

Thanks for your valuable comments. We hope our answers could help address the questions appropriately. `>>> C1` To our best understanding, the major goal of distributing submodels to clients is to train a shared and accurate global model collaboratively in a “top-down” manner. Such approaches are efficient for large-scale training, but they require that all the client models must be extracted from the same global model. In our setting, we focus on the “bottom-up” scenario to train a personalized model for each client. In other words, our goal is different from that of the “top-down” manner. In such a way, we do not need to require all the client's models to be extracted from the same global model, and all clients can have different network structures. Hence, the strict requirement of client models being extracted from a shared large model might be somewhat impractical in real-world applications, especially when compared to our model's design. It is important to acknowledge the efficiency and effectiveness of submodel-based federated learning models in large-scale training, particularly in scenarios where a shared global model is desired. `>>> C2` In the case of specific datasets like CIFAR-10, the data quantity remains constant. As the number of clients increases, the amount of distributed data per client decreases. Although federated learning (FL) iterates over all available data, clients working with limited datasets might not achieve the same level of performance as those with ample training data. Consequently, this dynamic can lead to a reduction in the overall performance. Recognizing the limitations, we acknowledge that our existing design might not be ideally suited for expansive cross-device federated learning scenarios. Nevertheless, it does exhibit effectiveness within the cross-silo setting. In our forthcoming version, we will explicitly define and delineate this cross-silo setting, along with an open discussion of the scalability challenges inherent in our current model. Your feedback significantly contributes to the refinement of our work.

Reviewer xdZD2023-08-10

I thank the authors for their answers. I have still some questions/remarks about the issues I mentioned. * the method does not seem suited for large-scale cross-device settings where clients may be selected only a few times. Perhaps my original remark was not clear. Consider a google-like cross-device setting where clients only participate a single time; your method would not work because no knowledge transfer from \hat w_t^n would be possible (client n never comes back). This is not the case for the personalized FL algorithms I mentioned in my review. Even if the client participates a few times, your algorithm should show significant performance deterioration. The 1/20 sampling rate is still very far from typical sampling rates in large scale cross-device settings. Perhaps the authors should acknowledge that their method is suited for cross-silo federated learning. * applicability of privacy-preserving techniques like secure aggregation. I understand the authors confirm that existing privacy-preserving techniques cannot be applied to their algorithm. * Computational overhead. Kairouz et al, Advances and Open Problems in Federated Learning, mentions that typical values of B are between 50 and 50000. What are typical values of M and what the size of stitching parts? * The candidate model can be more complex than the client's model. What are the constraints which make the candidate model almost always smaller than the client's model? * improvement w.r.t. state of the art. The new experiments confirm that one cannot conclude the improvement to be statistically significant for MNIST and SVHN.

Authorsrebuttal2023-08-12

Thank the Reviewer for the Constructive Comments

Dear reviewer, We do sincerely appreciate your constructive comments and questions. With all the respect, we provide our replies and hope they would address your concerns appropriately. `>>> Cross-device setting V.S. cross-silo setting` `>>> Answer:` Thanks for the insightful comments. The current model design is indeed not optimally suitable to the large-scale cross-device federated learning. Your invaluable input has enabled us to accurately define the scope of our contributions within the realm of cross-silo federated learning. In our final version, we will provide a comprehensive clarification of the cross-silo federated learning setting and expound upon the limitations of our work concerning the cross-device scenario. Your feedback greatly assists in enhancing the clarity and context of our research. `>>> Computational overhead` `>>> Answer:` Thanks for pointing out. In our current experimental setting, we set B as 10 when the number of clients is 100, aligning with the setting of existing work [1,2]. We can increase the value of B in our experiments, but it could lead to an increase of computational overhead on the server side, which is a limitation of our current model design. In our future work, we will explore new strategies to tackle this limitation and make the technique of heterogeneous model reassembly more efficient. Your insights serve as valuable guidance in refining our model. In response to your query about the parameter M, we would like to clarify that it is indeed a learned value in our model. In our experiments on the SVHN dataset with 100 clients and 0.1 active client ratio, we observed a dynamic range for M across multiple runs, spanning from 7 to 12. Regarding the size of stitching parts, we appreciate your interest in this aspect. As exemplified in Figure 4 of our original paper, we presented two generated candidate models. These models featured 4 stitching parts at time t and 5 stitching parts at time t', each corresponding to different depths. For each stitching part between two consecutive layers in the candidate model, we have a nonlinear activation function on top of a linear layer (sec 3.1.3). Notably, each stitching part involves only two parameters: $\mathbf{W}$ and $\mathbf{b}$, which dynamically adjust in dimensions based on the specific layers being stitched. Importantly, when compared to the parameter quantities present in the original models, the size of these parameters remains relatively modest. Furthermore, we only tune the stitching parts with fixing all the other parameters in the candidate models at the server side, which allows us to effectively enhance model performance while maintaining computational efficiency. [1] McMahan, Brendan, et al. "Communication-efficient learning of deep networks from decentralized data." Artificial intelligence and statistics. PMLR, 2017. [2] T Dinh, Canh, Nguyen Tran, and Josh Nguyen. "Personalized federated learning with moreau envelopes." Advances in Neural Information Processing Systems 33 (2020): 21394-21405. `>>> Improvement of state-of-the-art` `>>> Answer:`We appreciate your comments. It's worth noting that in comparison to the CIFAR-10 dataset, the MNIST and SVHN datasets are relatively less complex. While the performance enhancements on these latter two datasets may not always yield statistically significant results, it's still possible to discern discernible improvements. We would like to emphasize that these outcomes are based on scenarios where both client data and public data are derived from the same dataset. Furthermore, we have conducted another experiment to specifically investigate the impact of public data (Figure 1 in the original paper). Interestingly, when the public dataset and client data originate from different datasets, the performance gap between the proposed model and the baselines is further magnified. This underscores the significance of our approach and its potential for even greater differentiation in such scenarios.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC