Efficient Federated Learning against Heterogeneous and Non-stationary Client Unavailability

Addressing intermittent client availability is critical for the real-world deployment of federated learning algorithms. Most prior work either overlooks the potential non-stationarity in the dynamics of client unavailability or requires substantial memory/computation overhead. We study federated learning in the presence of heterogeneous and non-stationary client availability, which may occur when the deployment environments are uncertain, or the clients are mobile. The impacts of heterogeneity and non-stationarity on client unavailability can be significant, as we illustrate using FedAvg, the most widely adopted federated learning algorithm. We propose FedAPM, which includes novel algorithmic structures that (i) compensate for missed computations due to unavailability with only $O(1)$ additional memory and computation with respect to standard FedAvg, and (ii) evenly diffuse local updates within the federated learning system through implicit gossiping, despite being agnostic to non-stationary dynamics. We show that FedAPM converges to a stationary point of even non-convex objectives while achieving the desired linear speedup property. We corroborate our analysis with numerical experiments over diversified client unavailability dynamics on real-world data sets.

Paper

Similar papers

Peer review

Reviewer ZjRF6/10 · confidence 3/52024-07-12

Summary

The work studies heterogeneous and non-stationary client availability in FL, showing that it can have a significant impact on the performance of FedAvg. As a solution, the work presents the FedAWE algorithm, which compensate the missed computationd due to unavailability with minimal additional overhead w.r.t. FedAvg and is agnostic to non-stationary participation dynamics. The works presents a convergence proof in non-convex setting and numerical experiments to showcase the problem and assess the effectiveness of the proposed solution.

Strengths

1. **Importance of the research problem.** The work tackles an important and often overlook problem, that is heterogeneous and non-stationary client participation patterns. The research problem is particularly important given that in FL clients are edge devices, which are heterogeneous in computing capabilities and so availability patterns. 2. **Simple, principled and convenient approach.** The proposed approach is simple to implement following the details in algorithm 1, since it involves the calculation of a discout factor on clients' pseudo gradient and postponing the sending of the updated model. 3. **Good results.** The provided experimental results underscore the effectiveness of the approach in addressing both non-stationary and heterogeneous client availability. A plus is that authors did provide the code.

Weaknesses

The biggest weakness of the work is in my opinion the presentation. I found the writing not clear and straightforward enough to make the reader grasp the intuitions behind the technical innovations of the proposed approach. There are too many pointer to the supplementary material, and in most cases it seems matter that needs to be in main paper (e.g. line 131). Also, the implications of the technical lemmas are often poorly discussed (e.g. proposition 1-2, lemma 1-3), leaving the reader confused about the need of that particulare piece of information. I kindly suggest authors to revise the writing of the manuscript to better highlight the most imporant information to be conveyed, and avoid as much as possible continous pointer to the appendix.

Questions

In section 7, heterogeneous clients' data distribution are simulated $\nu_i \sim \text{Dirichlet(0.1)}$, as it is common in FL literature. Then the authors simulate the availabilty function of client $i$ as $p_i^t=p_i\cdot f_i(t)$, where $f_i(t)$ models the non-stationary dynamics and $p_i=\langle \nu_i, \phi\rangle$, with $\phi$ caractering the importance of each image class. Why $p_i$ is defined from the image class distribution $\nu_i$, given that it models the probability of being sampled? I expect the data distribution and the the availability to be uncorrelated, so maybe have $\mathbf{p} \sim \text{Dirichlet}(\gamma)$, where $\gamma$ decides how heterogeneous are the sampling probability of each client.

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

Limitations of the approach are not sufficiently discussed in section 8.

Reviewer ZjRF2024-08-08

Thanks to the authors for the elucidating response. I have also read and considered all the other reviews and relative authors' rebuttals. Since the provided information allowed me to better understand the work, I can more confidently confirm that the work provides interesting insights, and so I raised my confidence score accordingly. The authors promised to enhance the presentation, and based on the clear rebuttal I believe this will enhance the paper quality.

Authorsrebuttal2024-08-08

Thanks for the response

Dear Reviewer ZjRF, Thank you so much for increasing your confidence score! We are glad that you find our response to be elucidating and allowing you to better understand our work. We appreciate it and will further improve our presentation in the camera-ready version if accepted. Best regards, Authors of Submission 449

Reviewer i6Qf4/10 · confidence 4/52024-07-12

Summary

To address intermittent client availability, the authors study heterogeneous and non-stationary client availability, highlighting the significant impact of such heterogeneity using FedAvg. They propose FedAWE, which (i) compensates for missed computations due to unavailability and (ii) evenly diffuses local updates through implicit gossiping, despite non-stationary dynamics. The authors demonstrate that FedAWE converges to a stationary point for non-convex objectives while achieving the desired linear speedup property, supported by numerical experiments on real-world datasets with varied client unavailability dynamics.

Strengths

1) The authors propose a new algorithm to address the issue of intermittent client availability. 2) The paper is easy to follow, and a large number of cases w.r.t. unavailable dynamics are used for validation.

Weaknesses

1) The proposed algorithm is intuitive. The weight of the model update $G_i^t$ for each client is modified such that the longer the participation interval rounds, the greater the proportion of model update, which is proportional to $t-\tau(t)$. However, when the client participation rate is very low (e.g., 1% of clients participate randomly), $t-\tau(t)$will be expected to be 100, magnified by a factor of 100. This is somewhat unreasonable. 2) The rate of convergence in corollary1, the last term on the right-hand side, apparently dose not show a linear speedup with respect to the number of local steps $s$ and the number of active clients $k$. 3) In the rate of convergence in Corollary 1, it appears that $(1 + L^2)$ combines two variables with different "physical dimensions", which suggests that there might be an issue with the proof. For example, the "physical dimension" of the learning rate $\eta$ is $1/L$, and $(1 +\eta^2 L^2)$ is reasonable, while $(1 + L^2)$ might be incorrect. 4) It is not reasonable to see the performance of "FedAvg over all clients" being much lower than "FedAvg over active clients" in Table 1.

Questions

Can the authors provide an in-depth analysis on the algorithm's complexity?

Rating

4

Confidence

4

Soundness

2

Presentation

2

Contribution

2

Limitations

The authors have discussed the limitations of algorithms.

Reviewer DLNT7/10 · confidence 4/52024-07-25

Summary

This paper primarily focuses on addressing the issue of intermittent client availability in federated learning, where the problem scenario involves heterogeneity in participation and non-stationary client availability.The authors draw on ideas from other federated learning algorithms, including asynchronous federated learning, and introduce two novel algorithmic structures, "Adaptive Innovation Echoing" and "Implicit Gossiping", to address the issues of heterogeneity and non-stationary client unavailability.The authors demonstrate through experiments that heterogeneity and non-stationarity significantly impact the convergence performance of federated learning algorithms. Their newly proposed algorithm, FedAWE, exhibits the property of linear speedup while also saving substantial memory. The experimental results are comprehensive and show that their algorithm outperforms others, highlighting its advantages.

Strengths

The author's article combines the ideas of federated learning algorithms including asynchronous federated learning and peer-to-peer networks. Two novel algorithm architectures, "Adaptive Innovation Echoing" and "Implicit Gossiping", are proposed, which can greatly reduce memory. At the same time, the article is more general for client availability, only needs to satisfy independent and strictly positive across clients and rounds.

Weaknesses

In the section "Heterogeneity and Non-stationarity May Lead to Significant Bias", the authors only provide experimental evidence to verify that these factors affect the convergence performance of the algorithm. Could authors perhaps reference Theorem 1 and Theorem 2 from the FedAU paper for a theoretical analysis of these two factors? I believe that Theorem 1 from the article "A Lightweight Method for Tackling Unknown Participation Probabilities in Federated Averaging" could encompass the authors' experimental results concerning heterogeneity, that is, if the client participation is heterogeneous and the weight selection is not appropriate, it will directly affect the effect of convergence.

Questions

Firstly, do you believe that the conclusions regarding the heterogeneity part of your experiments in the section "Heterogeneity and Non-stationarity May Lead to Significant Bias" overlap with Theorem 1 from the paper "A Lightweight Method for Tackling Unknown Participation Probabilities in Federated Averaging"? Or do your conclusions offer any particular insights beyond this theorem? Secondly, the paper “FedVARP: Tackling the Variance Due to Partial Client Participation in Federated Learning” introduces the FedVARP algorithm, which is similar to the MIFA algorithm but employs a different variance reduction method. Could you also conduct related experiments to compare your algorithm with FedVARP?

Rating

7

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

The authors' work generalizes the application scenarios for federated learning by relaxing assumptions about client participation. It is hoped that in future research, the authors will incorporate variance reduction and other related techniques within the current algorithm framework to further enhance performance.

Reviewer DLNT2024-08-10

Thanks to the authors for the elucidating response.With the information you provided, I can better understand the significance of this paper, and the supplementary experiments also show the advantages of the algorithm.I believe that the author's efforts can enhance the overall quality of the paper.

Authorsrebuttal2024-08-10

Thanks for the response

Dear Reviewer DLNT, Thank you for confirming our efforts to enhance the overall quality of the paper. We greatly appreciate it! Best regards, Authors of Submission 449

Reviewer i6Qf2024-08-13

Further clarifications needed

I would like to thank the authors for providing the response. But I still have concerns regarding the gradient amplification and the appearance of $(1+L^2)$. 1) Gradient amplification: First, the authors provide some references, but these references are vague and offer limited relevant information on this specific point. For instance, [19] is a review, while [27, 52] focus on the convergence analysis of FedAvg. Second, in more extreme cases where the client participation rate drops to 1 in 1,000 or even 1 in 10,000, the gradient would be amplified by a factor of one thousand or ten thousand, which seems clearly unreasonable. The authors merely assert that their experiment is valid; however, they do not discuss or test this specific issue in their experimental analysis. 2) Appearance of $(1+L^2)$: More importantly, the authors acknowledge that the term $(1+\eta^2L^2)$ is dimensionally unified. Instead, they attempt to justify the reasonableness of the $(1+L^2)$ term by citing other references. However, upon reviewing the provided literature [27, 53, R1], I found no terms identical or even similar to $(1+L^2)$. Therefore, if this issue is not resolved, I do not think this paper is in a good shape for publication.

Authorsrebuttal2024-08-14

Thanks for the response

Dear Reviewer i6Qf: Thanks for your response. We apologize for any lack of clarity in the previous response due to insufficient space and hope that our further elaborations can address your concerns. **On gradient amplification.** - First, we are happy to elaborate the details of the cited references to corroborate our claim on gradient amplification. - [53] tackles stationary client unavailability, where the accumulated local gradients of each client are amplified by a factor obtained from a separate online estimation scheme. - [40] generalizes FedAvg in Algorithm 1 by amplifying the accumulation of local gradients by their corresponding responsive probabilities. - [19], as a classic survey in the federated learning community, points out open challenges in mitigating bias. For example, Section 7.2.3 on page 84 in [arXiv:1912.04977] mentions in the second bullet point that ''If the expected rate of contribution depends on factors outside our control, ..., one can correct for bias by scaling up or down contributions from devices depending on their stratum.'' - [52] studies non-stochastic client unavailability and proposes a generalized FedAvg algorithm, which is different from the vanilla FedAvg. In line 12 of its pseudocode, the accumulation of local gradients is amplified by a factor of $(\eta-1)$ every $P$ rounds. - [27], although studies partial device participation with known probability in Section 3.3, does not touch gradient amplifications. We are willing to take this off from the response. - Second, we are running the additional experiments under extremely small probabilities per your requests. We will update the experimental results in a separate response or, in the worst case, in the camera-ready version if accepted, as we are only less than one day away from the end of the author-reviewer discussion period. - In our submission, the contributing weight of each client's local gradient accumulation is not just $(t-\tau\_i(t))$ but a coupling of $\eta\_g(t-\tau\_i(t))$ due to a global learning rate $\eta\_g$ in line 10 of Algorithm 1. - For empirical success, the negative impacts of extremely long unavailable duration can be potentially mitigated through the carefully tuned global learning rate $\eta\_g$. Alternatively, we conjecture that clipping the unavailable duration up to a threshold like in FedAU [53] might help. It is also interesting to explore this direction theoretically in the future. **On the appearance of $(1+L^2)$.** After an additional careful review of our proofs, we did not identify the issues that will affect our presentations in Theorem 1 and Corollary 1. However, we discover an interesting fact that allows us to remove $L^2$ by following the same proof outline and learning rate conditions in $(10)$ but slightly different simplification techniques during lines 789 and 792 on page 34 and line 807 on page 37. The results hold in parallel to those in our initial submission, and we will incorporate the complete results, together with their proofs, into the camera-ready version if accepted. For ease of presentation, we discuss only the asymptotics w.r.t. $\eta\_l$ and $L$. The term $1/(mT)\sum\_{t=0}^{T-1}\sum\_{i=1}^m\mathbb{E}[\|\|\nabla F\_i(z\_i^t)\|\|^2]$ in lines 789, 792 and 806 can be expanded to three terms relating to $\zeta^2$, $\mathbb{E}[\|\|\nabla F(\bar{z}^t)\|\|^2]$ and $\mathbb{E}[\|\|z\_i^t-\bar{z}^t\|\|^2]$ by Proposition 3. Since $\eta\_l$ and $L$ share the same order in the first two, we are interested only in the last consensus error term, whose coefficient is proportional to $\eta\_l^3L^4$ ($\eta\_l^2L^4$ for line 807 since no rearrangement is needed). - *Sketch of the new derivation.* Directly plugging Lemma 6 on page 29 into lines 789, 792 to bound the consensus error term will yield the coefficients of the resulting terms in the scale of $\eta\_l^5L^4$. During the rearrangement in line 799, the order of $\eta\_l$ reduces by 1 and thus becomes $\eta\_l^4L^4$. Line 807 can be shown analogously. By choosing $\eta\_l=1/(sL\sqrt{T})$ for sufficiently large $T$, we will remove $L^2$. - *Sketch of the original derivation.* First, we apply the learning rate condition in (10) to simplify the coefficients of the respective terms in the consensus error bound (Lemma 6) on page 29. Then, the simplified consensus error bound is plugged into lines 789, 792 and 806. The resulting coefficients remain proportional to $\eta\_l^3L^4$ ($\eta\_l^2L^4$ for line 807). Therefore, $L^2$ remains after the rearrangement (one-order reduction) and the choice of learning rate. We appreciate the reviewer's efforts to check the provided literature in detail. We believe that this was only due to miscommunications, as we try to justify the appearance of $L$ instead of a union of $(1+L^2)$. Best regards, Authors of Submission 449

Reviewer i6Qf2024-08-14

Thanks for the response

Thank you for your detailed response. I believed the concern regarding gradient amplification can be addressed. Regarding the appearance of $(1+L^2)$, the authors have made efforts to resolve this by revising the proof, and I hope it can be fixed.

Authorsrebuttal2024-08-14

Thanks again for the response

Dear Reviewer i6Qf, Thank you for your confirmation. We appreciate it. Best regards, Authors of Submission 449

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC