A Kernel Perspective on Distillation-based Collaborative Learning

Over the past decade, there is a growing interest in collaborative learning that can enhance AI models of multiple parties. However, it is still challenging to enhance performance them without sharing private data and models from individual parties. One recent promising approach is to develop distillation-based algorithms that exploit unlabeled public data but the results are still unsatisfactory in both theory and practice. To tackle this problem, we rigorously analyze a representative distillation-based algorithm in the view of kernel regression. This work provides the first theoretical results to prove the (nearly) minimax optimality of the nonparametric collaborative learning algorithm that does not directly share local data or models in massively distributed statistically heterogeneous environments. Inspired by our theoretical results, we also propose a practical distillation-based collaborative learning algorithm based on neural network architecture. Our algorithm successfully bridges the gap between our theoretical assumptions and practical settings with neural networks through feature kernel matching. We simulate various regression tasks to verify our theory and demonstrate the practical feasibility of our proposed algorithm.

Paper

Similar papers

Peer review

Reviewer Zad85/10 · confidence 4/52024-07-08

Summary

This work presents an analysis of a distillation-based collaborative learning algorithm called FedMD from a nonparametric perspective. The method is to adopt an operator theoretic approach to obtain an upper rate for the expected generalization error of local models. Then, the authors propose DCL-KR to achieve a privacy-preserving nearly minimax optimal collaborative learning algorithm with kernel regression in massively distributed statistically heterogeneous environments.

Strengths

+ The idea seems to be interesting. + The first work tries to prove the (nearly) minimax optimality of a privacy-preserving collaborative learning algorithm.

Weaknesses

- The paper's organization and structure are not clear. - The writing looks like combining different parts of different papers without a clear clue. - The paper suddenly gives many assumptions in Section 3 for analysis but does not give any intuition or insight into them. - Some important experimental results are in the appendix. I would suggest that the authors extract insight into the experimental findings in the main body. - The language used in the paper is not concise.

Questions

The related work only gives a rough summary of the previous works. When the authors focus on the theoretical analysis, could the authors compare their theoretical bounds to other works? There are many important experiments in the appendix, such as C.3.1. Can the authors summarize empirical findings with a better paper structure? In Section 3.3, the paper starts with "To derive theoretical results" and lists a lot of assumptions without any intuitive understanding behind them. Could the authors give an overview and intuition of theoretical results? Minor: I feel very confused when reading the paper as it introduces many conclusions/results without too much explanation. I suggest the authors elaborate on the intuitions before going into the details.

Rating

5

Confidence

4

Soundness

2

Presentation

1

Contribution

3

Limitations

The paper presents a discussion of limitations, which is not convincing. For example, "there is no rigorous study that discusses the privacy preservation advantages of distillation-based collaborative learning". Actually, the standard technique for privacy is not distillation.

Reviewer UpcW6/10 · confidence 3/52024-07-09

Summary

The authors propose a nonparametric version of FedMD, a distillation-based collaborative learning methodology. They also propose a neural network variant of the nonparametric approach as an extension for heterogeneous local neural network models. They provide both theoretical results on the nonparametric approach and experimental results demonstrating efficacy of the neural network extension. The theoretical results prove the nearly minimax optimality of this nonparametric collaborative learning algorithm in massively distributed statistically heterogeneous environments.

Strengths

(1) The authors prove the (nearly) minimax optimality of a privacy-preserving collaborative learning algorithm using kernel regression in massively distributed, statistically heterogeneous environments. (2) Theorem 3.4 addresses a more general setting than prior works [48, 58]. Su et al. [58] only cover r = \frac{1}{2}​ of Assumption 3.3. Park et al. [48] do not consider Assumption 3.2, which provides a finer result. Furthermore, compared to [48], the required size of public inputs is reduced and the statistical homogeneity condition is dropped. (3) The rate n^{-\frac{r}{2r+s}} is the minimax lower rate under Assumptions 3.1, 3.2, and 3.3, making DCL-KR nearly optimal in a minimax sense. (4) On the difference between public and local distributions: Theorem 3.4 allows the public input distribution \tilde{\rho}_x to differ from the local input distribution \rho_x​. (5) Finally, The difference between \rho_x​ and \tilde{\rho}_x​ impacts the expected risk upper bound as a multiplication of B_r​, which can be removed by increasing public inputs.

Weaknesses

(1) The need to perform learning rate scaling to ensure the impact of local iterations is consistent across models introduces additional complexity. The exact method for scaling may vary and might not be trivial to determine or implement. (2) The computation of Gram matrices for CKA involves O(p^2) operations, where p is the number of public data points. This quadratic complexity can become prohibitive as the size of the public dataset increases. (3) The method assumes that the ensemble kernel k = \sum_{i=1}^{m} \frac{n_i}{n} k_{f_i} will perform better than individual feature kernels. This assumption might not hold in cases where local models are highly heterogeneous, as the averaged kernel could lose important characteristics of individual models, leading to suboptimal performance. The authors empirically verify the superiority of the ensemble in the appendix.

Questions

(1) How do you handle the computational burden associated with the calculation of Gram matrices for large datasets? (2) How do you handle cases where these representations before the last layers of the neural networks are suitable for kernel matching?

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

N/A

Reviewer UU4M6/10 · confidence 3/52024-07-11

Summary

The paper investigates the theoretical underpinnings and practical implementation of distillation-based collaborative learning (DCL) from a kernel regression perspective. The authors propose DCL-KR, a nonparametric version of the FedMD algorithm, which achieves nearly minimax optimal convergence rates in massively distributed and statistically heterogeneous environments. Building on these theoretical insights, the authors introduce DCL-NN, a practical DCL algorithm designed for neural networks, which incorporates feature kernel matching to align local models. Extensive experiments on various regression tasks demonstrate the superiority of DCL-KR and DCL-NN over existing methods.

Strengths

1. The paper provides a rigorous theoretical analysis of DCL-KR, proving its nearly minimax optimality in distributed and heterogeneous settings. 2. Experiments are comprehensive. Six datasets are included. 3. The improvement of DCL-NN is significant over the other baselines in some datasets.

Weaknesses

1. The paper lacks a discussion on the privacy risk and communication efficiency of DCL-KR and DCL-NN. 2. The paper is based on distillation-based collaborative learning method that requires public data, which limits the applications on private domains.

Questions

1. How does the choice of kernel affect the performance/theoretical guarantee of the proposed approach? 2. Is transferring predictions a common practice in collaborative learning for regression? If not, how does it compare with the other approaches in terms of privacy and efficiency?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes.

Reviewer 1rrx5/10 · confidence 1/52024-07-12

Summary

This paper theoretically proves a nonparametric version of the most standard distillation based collaborative learning algorithm (named DCL-KR) is nearly minimax optimal in massively distributed statistically heterogeneous environments. Extensive experiments demonstrate their theoretical results and show the practical feasibility of DCL-NN.

Strengths

1. The paper is well-written. 2. The authors conducted extensive experiments to verify the effectiveness of the proposed method.

Weaknesses

Please see Limitations

Questions

Please see Limitations

Rating

5

Confidence

1

Soundness

2

Presentation

2

Contribution

2

Limitations

This is not my field of study and I can only give a Borderline Reject. I will adjust my score based on the scores of other more specialised judges.

Reviewer 1rrx2024-08-13

After reading feedback from other reviewers, I am willing to improve my rating based on trusting the authors to modify the paper.

Reviewer Dtec7/10 · confidence 1/52024-07-15

Summary

In this paper, the authors perform a study of distillation-based collaborative learning (DCL) in massively distributed statistically heterogeneous environments. In particular, the study focuses on analyzing DCL-KR, a non-parametric version of FedMD, and proves its near-minimax optimality. The authors proposed DCL-NN, a method that leverages kernel matching to implement DCL with heterogeneous neural networks. Experimental results support the theoretical analysis of DCL-KR and demonstrate the effectiveness of DCL-NN.

Strengths

The paper provides a theoretical analysis of DCL-KR and proposes a practical algorithm for using neural networks in massively distributed statistically heterogeneous environments. The authors perform experiments with both synthetic and real-world datasets to support the theoretical results and evaluate the performance of DCL-NN.

Weaknesses

The paper does not perform any formal privacy analysis. As such, I believe using terms like "privacy-preserving" can be misleading in this context.

Questions

Could the authors avoid mentioning privacy as no formal privacy study is conducted?

Rating

7

Confidence

1

Soundness

4

Presentation

4

Contribution

4

Limitations

The paper unnecessarily emphasizes privacy while admitting that privacy is not formally defined (cf. question). It would be nice to provide more discussion on this aspect or to present some empirical evidence of the claimed benefit (e.g., performing privacy attacks such as membership inference)

Reviewer Dtec2024-08-11

Thank you for the responses

Thank you for the responses. I will keep my positive score.

Reviewer UU4M2024-08-09

Thanks for the author's response. The response has addressed my questions and I'll keep my positive score.

Reviewer 1rrx2024-08-09

Thank you for the author's response. I will take into account the feedback from other reviewers to adjust my scoring subsequently.

Reviewer Zad82024-08-09

Thanks for the detailed explanation and response. After reading the whole page, I am willing to improve my rating based on trusting the authors to modify the paper. Thank you very much.

Reviewer UpcW2024-08-13

Thanks for the response. This has cleared up most concerns for me. On this comment, >Therefore, with enough data, it is not possible for the ensemble kernel to perform worse than individual local feature kernels. However, fully distilling the ensemble kernel is challenging. Thus, there is a possibility of losing important characteristics (that the reviewer concerned with) during the kernel distillation process. Precisely, some important features may have a lower portion in the ensemble kernel and so they can be potentially missed during the distillation process. One possible solution is for each local party to construct local feature kernels by selecting important features and adjusting weights accordingly. This is particularly what I was aiming at. On re-reading my review, I realise I was not clear. I agree with enough data the ensembled kernel cannot be worse than the local kernels. I agree the problem here is the distillation process and that piece should be explored in future work. There is work in leveraging bochner's theorem for learning kernels from data in the gaussian process literature. Perhaps these approaches could provide more optimal kernels at the local level. I've updated my score given this response.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC