Summary
This paper proposes a new FL setting that involves a dynamically moving server among clients residing in different regions and receiving updates from clients of the selected region through wireless links. The proposed algorithm RWSADMM addresses personalization and data heterogeneity by minimizing the local proximity among neighbor nodes of each client. Additionally, the authors apply a first-order subgradient expansion for computational complexity reduction, which is useful for real-world applications. The algorithm is demonstrated to converge provably faster than baseline methods through theoretical analysis and experimental comparisons.
Strengths
This paper introduces a novel asynchronous mobilizing federated learning setting that can be
utilized in realistic war situations where communication infrastructure is unavailable. Additionally,
it establishes a mathematical framework based on dynamic graph-based representation, allowing the server to move through the field and interact with various clients to update the neural network.
The proposed algorithm surpasses the baseline methods by a significant margin, demonstrating notably faster convergence rates, particularly in realistic non-IID scenarios.
The methodology employed in this study is a theoretically sound solution derived from the analysis of dynamic Markov Chain and Alternating Direction Method of Multipliers.
The proposed RWSADMM algorithm ensures convergence based on commonly assumed properties of the loss function. Furthermore, the communication complexity of RWSADMM is compared to existing methods like Per-FedAvg and APFL, providing theoretical evidence of its superior scalability.
In the proposed setting, both convex and non-convex models converge faster and exhibit higher accuracy performance compared to the baseline methods.
Weaknesses
The proposed setting (mobilized federated learning) seems somewhat unrealistic. In a war situation where long-range communication is not supported, is there a need for federated learning? Considering that neural networks are large and transmitting them via tactical vehicles would be challenging, wouldn't it be better to transmit data for distributed learning or perform centralized learning on the server, rather than transmitting model updates?
The proposed Markov Chain does not seem to effectively represent the problem situation (mobilized federated learning). In real scenarios, nodes within the communication range of the communication vehicle should be selected, but the proposed algorithm does not consider the physical distances between nodes. Additionally, considering realistic situations like Figure 3, where nodes (walking soldiers) move and the topology keeps changing, the Markov Chain used in this paper cannot reflect spatial rearrangements of participating nodes.
I am concerned about privacy issues. In the proposed algorithm, each node needs to receive model updates from neighboring nodes to calculate the proximity term. However, this process can expose the data that the node possesses, as it can be extracted from the updates of other nodes. Furthermore, simultaneous reception of model updates from multiple nodes can result in communication overhead.
There is a lack of ablation study on the proposed components. There is no analysis on how well the proposed local proximity term resolves data heterogeneity, and no trade-off analysis on how much computational resources are saved through first-order expansion and the corresponding impact on performance.
Questions
In the current problem, the server moves through a random walk. However, in practical situations, shouldn't the server actively modify its routing to reach unseen nodes?
Looking at the Experimental setup in Section 5, there are only 20 clients. Considering the trend in recently published Federated learning works, shouldn't the proposed algorithm be compatible with scenarios involving a larger number of clients?
In Figure 2, although PerAvg and APFL are FL works geared to data heterogeneity, their performances are significantly lower compared to FedAvg. Why does this discrepancy occur?
What are the confidence intervals for the reported performance in Figure 2 and Table 1? It is important to report confidence intervals, as performance can vary significantly depending on how the spatial relationship between nodes is initialized.
The proposed setting (mobilized federated learning) seems somewhat unrealistic. In a war situation where long-range communication is not supported, is there a need for federated learning? Considering that neural networks are large and transmitting them via tactical vehicles would be challenging, wouldn't it be better to transmit data for distributed learning or perform centralized learning on the server, rather than transmitting model updates?
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.