Free-Rider and Conflict Aware Collaboration Formation for Cross-Silo Federated Learning

Federated learning (FL) is a machine learning paradigm that allows multiple FL participants (FL-PTs) to collaborate on training models without sharing private data. Due to data heterogeneity, negative transfer may occur in the FL training process. This necessitates FL-PT selection based on their data complementarity. In cross-silo FL, organizations that engage in business activities are key sources of FL-PTs. The resulting FL ecosystem has two features: (i) self-interest, and (ii) competition among FL-PTs. This requires the desirable FL-PT selection strategy to simultaneously mitigate the problems of free riders and conflicts of interest among competitors. To this end, we propose an optimal FL collaboration formation strategy -- FedEgoists -- which ensures that: (1) a FL-PT can benefit from FL if and only if it benefits the FL ecosystem, and (2) a FL-PT will not contribute to its competitors or their supporters. It provides an efficient clustering solution to group FL-PTs into coalitions, ensuring that within each coalition, FL-PTs share the same interest. We theoretically prove that the FL-PT coalitions formed are optimal since no coalitions can collaborate together to improve the utility of any of their members. Extensive experiments on widely adopted benchmark datasets demonstrate the effectiveness of FedEgoists compared to nine state-of-the-art baseline methods, and its ability to establish efficient collaborative networks in cross-silos FL with FL-PTs that engage in business activities.

Paper

Similar papers

Peer review

Reviewer 2xqy6/10 · confidence 4/52024-06-15

Summary

This paper introduces a strategy named FedEgoists designed to enhance collaboration in cross-silo federated learning (FL) scenarios, particularly in business sectors where participants (FL-PTs) are often competitive and self-interested. The proposed FedEgoists strategy presents a sophisticated and theoretically sound approach to managing collaborations in cross-silo federated learning, and its strengths lie in its ability to handle self-interest and competition effectively.

Strengths

FedEgoists ensures that the formed coalitions are optimal. No coalition can improve its utility by merging with another coalition, making the solution stable and efficient. The strategy effectively addresses the problem of free riders, ensuring that all FL-PTs contribute to and benefit from the FL ecosystem proportionately. By preventing FL-PTs from contributing to their competitors or their supporters, the strategy minimizes conflicts of interest. This paper validates FedEgoists using the benchmark experimental datasets.

Weaknesses

The paper lacks rigorness as pointed out in the Questions part.

Questions

(1) FedEgoists requires a central server to coordinate and enforce the coalition formation, which introduces a single point of control and potential failure. This centralization might conflict with the decentralized nature of federated learning and violate the fundamental motivation of FL. (2) FedEgoists assumes that FL-PTs will accurately report their competitive relationships to the central server (CS). However, this can easily lead to privacy leaks and potential attacks. (3) While the FedEgoists strategy focuses on collaboration benefits and competition avoidance, it does not explicitly address data complementarity. (4) Since the FedEgoists strategy relies on coalition formation based on reported benefits and competition, there is a risk of strategic manipulation by participants. (5) The experiments are not sufficient, and there are too few datasets and baselines.

Rating

6

Confidence

4

Soundness

2

Presentation

2

Contribution

2

Limitations

It seems limitations are not clearly or explictly elaborated.

Reviewer gR3Y6/10 · confidence 4/52024-07-05

Summary

This paper studies an interesting topic, which is about cooperation and competition in federated learning. It can be used to describe or simulate the real-world scenarios. The authors propose to use graph-related techniques to formulate the relation between the local clients. They test the algorithm on the benchmark datasets and the medical dataset.

Strengths

1, The research question in federated learning may have potential practical impact in the real-world setting. The cross-silo setting fits what the authors explore. Considering the cooperation and competition from the graph perspective in FL is an interesting point. 2, Statements are supported with experiment results and theory. 3, The presentation is helpful for readers including the algorithm description, diagram, and figures.

Weaknesses

1, The novelty of each module, especially the graph part, may need to be clarified clearly. Also, the two principles are commonly used, if the reviewer understand correctly, an optimal formation is proposed based on these two, is that correct? 2, Most designed part is at the server side. The reviewer may not be able to observe the strong connection with federated learning. The setting itself may tell that each client is a participant in this framework. However, the introduction (motivation) and the method could be a little bit fragmented. 3, Ablation study and hyperparameter study may help verify the algorithm more deeply and soundly. 4, A little bit concerned about the baseline selection. Some contribution-related work in FL like [1], if comparable, should be considered or at least discussed. [1] Contribution-Aware Federated Learning for Smart Healthcare

Questions

1, Could you please address the concerns in the weakness part? 2, Could you clarify how the local model information is leveraged and updated in the graph part more clearly? 3, How to validate or verify the meaning of the graph construction part? Do we need to have the groundtruth of the relation of all the clients?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

1, The connection of the proposed techniques should be closely associated with the setting of FL, which focuses on the iterative updates between the server and the clients. 2, The original novelty of contribution should be emphasized a little bit more. 3, More comprehensive experiments may help demonstrate the effectiveness and understand the generalization of the algorithm better.

Authorsrebuttal2024-08-07

**Weaknesses 4** The two papers have different focuses. There is a central server that coordinates the FL training process. In [1], a global ML model is built for all clients and the paper aims to evaluate the individual contribution of each client/FL-PT to this global model. In our paper, a personalized ML model is produced for each individual client $v_{i}$ and the central server determines which clients can contribute to this client $v_{i}$ in the FL training process. While determining such contribution relationships between clients, we aim to avoid free-riding and conflicts of interests between clients.

Authorsrebuttal2024-08-07

**Questions 2 \& 3** This paper has two graphs (i.e., the benefit graph $\mathcal{G_b}$, and the competing graph $\mathcal{G_c}$) that are assumed to be known to a FL manager; $\mathcal{G_b}$ is obtained by collaborative model training, while the information on $\mathcal{G_c}$ is reported by clients to the FL manager (or the central server). Below, we detail the ways of obtaining them and how the local model information is leveraged in the graph part. Firstly, data heterogeneity/complementarity between clients is characterized by a benefit graph $\mathcal{G_b}$. Like [5,30], this paper uses the hypernetwork technique in [23] to evaluate the data complementarity, thus obtaining the benefit graph. In the graph part, the local model information is mainly leveraged in the calculation of $\mathcal{G_b}$. We also refer readers to Section 4.2 of Ref. [5] for the technical details on the way of obtaining $\mathcal{G_b}$. Specifically, there are $n$ clients. Each client $v_{i}$ has a risk/loss function $\ell_{i}$: $\mathbb{R^n}\rightarrow \mathbb{R_+}$. Given a learned hypothesis $h\in$ $\mathcal{H}$, let the loss vector $\mathbf{\ell}(h)=[\ell_{1}, \dots, \ell_{n}]$ represent the utility loss of the $n$ clients under the hypothesis $h$. The hypothesis $h$ is considered a Pareto solution if there is no other hypothesis $h^{\prime}$ that dominates $h$, i.e., $\nexists h^{\prime}\in \mathcal{H}, s.\,t.\, \forall i: \ell_{i}(h^{\prime})\leqslant \ell_{i}(h) \text{ and } \exists j: \ell_{j}(h^{\prime}) < \ell_{j}(h). $ Let $r=(r_{1}, \dots,r_{n})$ $\in \mathbb{R^n}$ denote a preference vector which denotes the weight of the objective local model loss that is normalized with $\sum_{k=1}^{n}{r_{k}}=1$ and $r_{k}\geq 0, \forall k\in \{1, \dots, n\}$. The hypernetwork $HN$ takes $r$ as input and outputs a Pareto solution $h$, i.e., $h\gets HN(\phi, r)$, where $\phi$ denotes the parameters of the hypernetwork [23]. For each client $v_{i}$, linear scalarization can be used. Like [5,30], an optimal preference vector $r_{i}^{\ast}=\left(r_{i,1}^{\ast}, r_{i,2}^{\ast}, \dots, r_{i,n}^{\ast}\right)$ is determined to generate the hypothesis $h_{i}^{\ast}$ that minimizes the loss with the data $\hat{\mathcal{D_i}}$. This is expressed as $h_{i}^{\ast}=HN(\phi, r_{i}^{\ast}), \text{ where } r_i^* = argmin_r \hat{\mathcal{L_i}}(HN(\phi, r)).$ For each client $v_{i}$, the value of $r_{i,j}^{\ast}$ is used as an estimate to the weight of $v_{j}$ to $v_{i}$ [5,30]. $r_{1}^{\ast}, r_{2}^{\ast}, \dots, r_{n}^{\ast}$ define a directed weighted graph, i.e., the benefit graph $\mathcal{G_b}$. Secondly, following the standard technical practices in literature, the central server is assumed to be trustable, as explained above. In real worlds, the central server may represent an impartial and authoritative third-party (e.g., the industry association) [5]. Then, FL-PTs can report their competitive relationships to the third-party in person and confidentiality agreements can be signed between the third-party and FL-PTs. The above information will be clarified in the final version of our manuscript.

Authorsrebuttal2024-08-07

**Limitations 3** Due to space limitations, please refer to the global rebuttal.

Reviewer gR3Y2024-08-08

Thanks for the author's comments. I think the comments have addressed my concerns and answered the questions. I will increase the score.

Reviewer BcXD7/10 · confidence 5/52024-07-07

Summary

The business sector is a main domain where cross-silo federated learning (FL) has many promising applications in various scenarios. The authors simultaneously consider the self-interest and competition features in the business sector. They develop a novel framework to both address the resulting free-riding problem and avoid the conflict of interest between any two competing FL participants (FL-PTs), which results in an interesting optimization problem. Finally, the authors find an optimal solution where one coalition cannot increase the utility of any of its members by collaborating with other coalitions. Extensive experiments are conducted to show the effectiveness of the proposed solution, and its ability to establish efficient collaborative networks in cross-silo FL with FL-PTs that engage in business activities.

Strengths

S1: The authors consider a timely and interesting question in cross-silo FL where the free-riding and competition issues are considered simultaneously. S2: The authors propose a new problem formulation and definition and develop a novel theoretical framework of practical importance to address the identified problem. The resulting optimization problem is solved optimally. The paper is technically sound. S3: The paper is well organized and easy to follow. The background and related work are well introduced to understand the importance of the problem of this paper. S4: Extensive experiments have been done to show the effectiveness of the proposed solution.

Weaknesses

W1: The proposed algorithm is not polynomial-time solvable. However, this seems reasonable since there are typically a limited number of FL-PTs (e.g., 2 to 100) for cross-silo FL in business sectors. The authors propose a novel application of the classic algorithms in graph theory to solve their problem optimally. The algorithm complexity depends on these classic graph algorithms, which work well in reality.

Questions

Q1: Can the proposed technical framework be applied to other types of tasks or datasets, e.g., NLP? In the experimental evaluation part, the authors have already shown its application to some interesting tasks commonly seen in literature. Q2: To keep consistency, it is better that “selfish” in Figure 1 is changed to “self-interest”. Q3: In this paper, all FL-PTs are partitioned into mutually disjoint coalitions/groups. It seems that this is similar to clustered federated learning where FL-PTs are partitioned into mutually disjoint clusters (i.e., groups). The reviewer understands that the authors consider a problem that is clearly different from the typical clustered federated learning. Q4: In your experimental results, how many trials have you conducted to obtain the standard deviation?

Rating

7

Confidence

5

Soundness

3

Presentation

4

Contribution

3

Limitations

See weaknesses and Questions.

Reviewer jdA34/10 · confidence 4/52024-07-13

Summary

This paper focuses on client selection in cross-silo federated learning. The authors propose FedEgoists. In particular, FedEgoist participate clients into different clusters to avoid free riders and conflict of interests. Theoretical analysis is provided to validate the theoretical soundness of the proposed FedEgoist. Experiments validate the effectiveness of proposed FedEgoist.

Strengths

* The discussed topic, client selection in cross-silo federated learning, is important in practical setting. * The writing is easy to follow.

Weaknesses

* The evaluation setting on CIFAR-10 and CIFAR-100 cannot simulate conflict of interests in the real world. Further consideration needed. * The meaning of propositions 1 & 2 is unclear. Further explanation is expected.

Questions

Please refer to weaknesses.

Rating

4

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

There is no potential negative societal impact of their work.

Authorsrebuttal2024-08-07

**Notes on other added experiments.** In this rebuttal, besides the experiments conducted for the above comments in W1, more experiments are also conducted (a) to verify the robustness of the proposed solution in terms of the level of data heterogeneity, and (b) on a new dataset and with a new baseline approach. **Experiments (a)** Specifically, this paper is about the application of the current FL techniques to the business sectors, and the experiments aim to verify the effectiveness of the proposed theoretical framework in the scenario under study. The scenario features self-interest for every client and competition between a part of clients. The collaboration relationships between clients are built according to the competing and benefit graphs; the benefit graph depends on the data heterogeneity/complementarity of clients. Thus, the variable factors in the scenario/experimental environment include the intensity of competition between clients and the non-IID setting. In the experiments, we vary the related parameters in these two factors to better verify the robustness of the proposed algorithm. Above, the experiments for the intensity of competition have been introduced. Data heterogeneity is typically simulated by a pathological or Dirichlet distribution. For example, the pathological distribution is used in [5,30]. In this paper, we conducted experiments on pathological and Dirichlet distributions, respectively. Also, in this rebuttal, we better verify the effect of data heterogeneity on the effectiveness of the proposed solution. Specifically, let $m$ denote the number of classes. In Dirichlet distribution, there is a distribution vector $q_{c}\in \mathbb{R}^{m}$ is drawn from the Dirichlet distribution $Dir_{m}(\beta)$ for each class $c$ and client $v_{i}$ is allocated a $q_{c,i}$ proportion of data samples of class $c$; smaller $\beta$ value results in higher data heterogeneity. We vary the value of $\beta$ that takes different values in {0.01,0.1,0.5} and conducted the corresponding experiments. The related experimental results are presented in Table 4. **Experiments (b)** We additionally conducted experiments on the synthetic data used in [30] where the experimental setting is also the same as the setting in [30]. The related experimental results are presented in Tables 5-6, which can be found in the PDF document. We add one more approach in the paper below, called FEDORA, as the baseline: Jun Wu, Wenxuan Bao, Elizabeth Ainsworth, and Jingrui He. "Personalized federated learning with parameter propagation." In ACM KDD, 2023. FEDORA is only for image classification tasks; thus, we only conducted experiments on CIFAR-10 and CIFAR-100. The related experimental results are presented in Tables 1,2,4.

Authorsrebuttal2024-08-11

Discussion Period & Thank you!

Dear Reviewer jdA3, we thank you for your time and detailed comments. Since the discussion period will end soon, we will appreciate if you could check the rebuttal and let us know whether our response has addressed your concerns. Thank you sincerely.

Reviewer BcXD2024-08-12

Thanks for the detailed responses. I'm generally fine with the results. Thus, I will keep my score.

Authorsrebuttal2024-08-12

Thank you!

Dear Reviewer BcXD, we thank you sincerely for your time to check the rebuttal and your recognization of our work.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC