Self-Centered Federated Learning (SCFL): An Overview

• SCFL is a federated learning approach that shifts focus from a universal global model to client-specific objectives, including selfish utility maximization, personalization, and decentralized federation formation.
• It employs techniques such as robust median-based aggregation and personalized weighting to mitigate biased or adversarial client updates while preserving useful signal.
• Empirical studies on datasets like MNIST, CIFAR-10, and EMNIST demonstrate SCFL’s effectiveness in balancing individual client benefits with overall model convergence and performance.

Self-Centered Federated Learning (SCFL) denotes a family of federated learning formulations in which collaboration is centered on the utility, similarity structure, or collaboration scope of individual participants rather than on a single undifferentiated global model. In the recent arXiv literature, however, the acronym is not semantically unique. It is used for selfish clients that steer the shared model toward their own optimum, for attacker-centric schemes that simultaneously damage the global model and improve a private model, for server-side personalized aggregation, for decentralized proximity-based federation formation, and, in unrelated usages, for “Sample Clustered Federated Learning” and “Social-aware Clustered Federated Learning” (Augello et al., 2024, Zhang et al., 30 Aug 2025, Mestoukirdi et al., 2023, Mestoukirdi et al., 2021, Domini et al., 2024, Berdoz et al., 2022, Manthe et al., 2024, Wang et al., 2022).

1. Terminology and conceptual scope

The literature uses “SCFL” for several distinct mechanisms. In one line of work, self-centered behavior means that a participant attempts to center the global trajectory on its own data distribution. In another, it means that each client is the center of its own personalized aggregation or knowledge-distillation pipeline. In yet another, the acronym expands to notions that are not “Self-Centered” at all, such as “Sample Clustered Federated Learning” and “Social-aware Clustered Federated Learning” (Manthe et al., 2024, Wang et al., 2022).

Usage of SCFL	Core mechanism	Representative papers
Self-centered utility maximization	Clients or attackers increase their own influence or private advantage	(Augello et al., 2024, Zhang et al., 30 Aug 2025)
User-centric personalization	The server or client computes personalized collaboration rules	(Mestoukirdi et al., 2021, Mestoukirdi et al., 2023, Berdoz et al., 2022)
Proximity-based self-federation	Peer-to-peer federation formation without a central server	(Domini et al., 2024)
Other acronym expansions	Sample-level domain clustering or social-trust clustering	(Manthe et al., 2024, Wang et al., 2022)

A recurrent misconception is to treat SCFL as a single canonical algorithmic family. The cited works do not support that reading. A more accurate interpretation is that SCFL is an overloaded label applied to several client-centric departures from conventional FedAvg-style global averaging. This suggests that the term should be read only together with the local expansion and threat or system model adopted by a given paper.

2. Self-centered utility maximization: selfish clients and adversarial self-benefit

In “Tackling Selfish Clients in Federated Learning,” selfish clients are participants who deliberately deviate from standard training to make the global model inclined toward their local model, thereby prioritizing their local data distribution. They are described as rational and non-malicious: they do not intend to poison the model or break convergence; they aim to reduce their own loss on their local data by increasing their influence on the global update (Augello et al., 2024).

The baseline federated objective is

$$\mathbf{w}^* \;=\; \arg\min_{\mathbf{w}} \Big\{ F(\mathbf{w}) \triangleq \tfrac{1}{k} \sum_{i\in [k]} F_i(\mathbf{w}) \Big\},$$

with server update

$$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$

For a selfish client $s$, the manipulated update is chosen to minimize local post-aggregation loss: $$\hat{\delta}_s \;=\; \arg\min_{\delta} \; F_s\!\Big(\hat{\mathbf{w}}\Big), \quad \text{where} \quad \hat{\mathbf{w}} \;\triangleq\; \mathbf{w} + \frac{\delta_{[k]\setminus\{s\}} + \delta}{k}.$$ The paper introduces a selfishness parameter $\alpha\in[0,1]$ and biased rule

$$\hat{\delta}_s \;=\; \alpha\!\left(k\,\delta_s - (k-1)\,\bar{\delta}_{[k]\setminus\{s\}}\right) + (1-\alpha)\,\bar{\delta}_{[k]\setminus\{s\}} \;=\; \alpha k\big(\delta_s - \bar{\delta}_{[k]\setminus\{s\}}\big) + \bar{\delta}_{[k]\setminus\{s\}}.$$

Special cases are explicit: $\alpha=0$ gives no selfish effect, $\alpha=\tfrac{1}{k}$ gives the honest update, and $\alpha=1$ gives a full replacement trend. Experiments on MNIST and CIFAR-10 show that “Just 2% of clients behaving selfishly can decrease the accuracy by up to 36%,” and that multiple selfish clients can destabilize convergence; without mitigation, there is no convergence for $\ge 3$ selfish clients.

“FedThief: Harming Others to Benefit Oneself in Self-Centered Federated Learning” formalizes a more explicitly malicious SCFL regime. Here, adversaries degrade the global model while simultaneously improving a private model available only to themselves. The objective is quantified by the malicious advantage

$$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$0

where $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$1 is the final global model accuracy and $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$2 is the attacker’s private ensemble-model accuracy. FedThief uses four local models per malicious client: a private model $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$3, a malicious model $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$4, an error model $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$5, and an ensemble head $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$6. Upload-stage manipulation is

$$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$7

and the private objective is

$$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$8

The reported results show simultaneous global degradation and private improvement across datasets and defenses. For example, on CIFAR-10 with FedAvg, Min-Sum, and $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$9, $s$0, $s$1, and $s$2; on MNIST with Pair Flip, Bulyan, and $s$3, $s$4, $s$5, and $s$6 (Zhang et al., 30 Aug 2025).

Taken together, these two lines separate two technically different meanings of self-centeredness. In the selfish-client formulation, the client seeks extra influence while still valuing the common task. In the FedThief formulation, the attacker preserves the degradation caused by poisoning attacks while extracting private benefit. The distinction is central: one is rational, non-malicious deviation; the other is adversarial bi-objective optimization.

3. Robust aggregation against selfishness

The main server-side mitigation proposed for selfish self-centered behavior is RFL-Self, a drop-in replacement for FedAvg that detects suspected selfish clients via robust statistics on update norms and then recovers an estimate of their true update before aggregation (Augello et al., 2024).

Detection uses the median of $s$7 norms,

$s$8

and flags client $s$9 as suspected selfish if $$\hat{\delta}_s \;=\; \arg\min_{\delta} \; F_s\!\Big(\hat{\mathbf{w}}\Big), \quad \text{where} \quad \hat{\mathbf{w}} \;\triangleq\; \mathbf{w} + \frac{\delta_{[k]\setminus\{s\}} + \delta}{k}.$$0. The server also computes a robust central update vector

$$\hat{\delta}_s \;=\; \arg\min_{\delta} \; F_s\!\Big(\hat{\mathbf{w}}\Big), \quad \text{where} \quad \hat{\mathbf{w}} \;\triangleq\; \mathbf{w} + \frac{\delta_{[k]\setminus\{s\}} + \delta}{k}.$$1

namely the coordinate-wise median across clients. The paper proves a norm-inflation theorem: if a selfish client’s true update is similar in magnitude to the mean update of normal clients, then any effective selfish estimate $$\hat{\delta}_s \;=\; \arg\min_{\delta} \; F_s\!\Big(\hat{\mathbf{w}}\Big), \quad \text{where} \quad \hat{\mathbf{w}} \;\triangleq\; \mathbf{w} + \frac{\delta_{[k]\setminus\{s\}} + \delta}{k}.$$2 is necessarily larger in norm than the true update $$\hat{\delta}_s \;=\; \arg\min_{\delta} \; F_s\!\Big(\hat{\mathbf{w}}\Big), \quad \text{where} \quad \hat{\mathbf{w}} \;\triangleq\; \mathbf{w} + \frac{\delta_{[k]\setminus\{s\}} + \delta}{k}.$$3. This supplies principled support for the median-of-norms test.

Rather than dropping flagged updates, RFL-Self reconstructs them: $$\hat{\delta}_s \;=\; \arg\min_{\delta} \; F_s\!\Big(\hat{\mathbf{w}}\Big), \quad \text{where} \quad \hat{\mathbf{w}} \;\triangleq\; \mathbf{w} + \frac{\delta_{[k]\setminus\{s\}} + \delta}{k}.$$4 with $$\hat{\delta}_s \;=\; \arg\min_{\delta} \; F_s\!\Big(\hat{\mathbf{w}}\Big), \quad \text{where} \quad \hat{\mathbf{w}} \;\triangleq\; \mathbf{w} + \frac{\delta_{[k]\setminus\{s\}} + \delta}{k}.$$5 chosen to satisfy

$$\hat{\delta}_s \;=\; \arg\min_{\delta} \; F_s\!\Big(\hat{\mathbf{w}}\Big), \quad \text{where} \quad \hat{\mathbf{w}} \;\triangleq\; \mathbf{w} + \frac{\delta_{[k]\setminus\{s\}} + \delta}{k}.$$6

The final aggregation is

$$\hat{\delta}_s \;=\; \arg\min_{\delta} \; F_s\!\Big(\hat{\mathbf{w}}\Big), \quad \text{where} \quad \hat{\mathbf{w}} \;\triangleq\; \mathbf{w} + \frac{\delta_{[k]\setminus\{s\}} + \delta}{k}.$$7

Under the stated conditions, solving the norm equation yields $$\hat{\delta}_s \;=\; \arg\min_{\delta} \; F_s\!\Big(\hat{\mathbf{w}}\Big), \quad \text{where} \quad \hat{\mathbf{w}} \;\triangleq\; \mathbf{w} + \frac{\delta_{[k]\setminus\{s\}} + \delta}{k}.$$8, so $$\hat{\delta}_s \;=\; \arg\min_{\delta} \; F_s\!\Big(\hat{\mathbf{w}}\Big), \quad \text{where} \quad \hat{\mathbf{w}} \;\triangleq\; \mathbf{w} + \frac{\delta_{[k]\setminus\{s\}} + \delta}{k}.$$9 is close to the true $\alpha\in[0,1]$0. The bounded aggregation-error theorem gives

$\alpha\in[0,1]$1

as a bound on the maximum error in the recovered aggregated update, with insensitivity to the number of selfish clients as long as they do not substantially bias the median.

The empirical evaluation uses MNIST and CIFAR-10 with $\alpha\in[0,1]$2 clients, full participation, $\alpha\in[0,1]$3 local epochs, and 30 rounds. At $\alpha\in[0,1]$4 and 10% selfish clients on CIFAR-10, normal clients’ accuracy is 60.84% for RFL-Self, 56.64% for Downscaling, and 55.47% for Median; selfish clients’ accuracy is 56.00%, 54.20%, and 56.20%, respectively. Across selfish fractions $\alpha\in[0,1]$5 and $\alpha\in[0,1]$6, RFL-Self outperforms Median by 7–12% and Downscaling by 4–5% on MNIST for $\alpha\in[0,1]$7, and it does not degrade accuracy when there are 0% selfish clients. The paper also reports stability in a few-client regime with $\alpha\in[0,1]$8.

The design choice is notable. Byzantine defenses such as Krum, Multi-Krum, Bulyan, Trimmed Mean, Median, and RSA are described as primarily excluding or heavily downweighting nonconforming updates. RFL-Self instead tries to preserve useful signal from self-centered clients by estimating the unmanipulated update. This is a materially different robustness objective from standard adversarial filtering.

4. Personalized server-side SCFL: user-centric aggregation

A separate SCFL lineage treats self-centeredness as personalization. In “User-Centric Federated Learning” and “User-Centric Federated Learning: Trading off Wireless Resources for Personalization,” the parameter server computes a distinct aggregation rule for each client, so that client $\alpha\in[0,1]$9 receives a personalized model

$$\hat{\delta}_s \;=\; \alpha\!\left(k\,\delta_s - (k-1)\,\bar{\delta}_{[k]\setminus\{s\}}\right) + (1-\alpha)\,\bar{\delta}_{[k]\setminus\{s\}} \;=\; \alpha k\big(\delta_s - \bar{\delta}_{[k]\setminus\{s\}}\big) + \bar{\delta}_{[k]\setminus\{s\}}.$$0

Here, the collaboration vector $$\hat{\delta}_s \;=\; \alpha\!\left(k\,\delta_s - (k-1)\,\bar{\delta}_{[k]\setminus\{s\}}\right) + (1-\alpha)\,\bar{\delta}_{[k]\setminus\{s\}} \;=\; \alpha k\big(\delta_s - \bar{\delta}_{[k]\setminus\{s\}}\big) + \bar{\delta}_{[k]\setminus\{s\}}.$$1 is client-specific and is designed to optimize performance on the target distribution $$\hat{\delta}_s \;=\; \alpha\!\left(k\,\delta_s - (k-1)\,\bar{\delta}_{[k]\setminus\{s\}}\right) + (1-\alpha)\,\bar{\delta}_{[k]\setminus\{s\}} \;=\; \alpha k\big(\delta_s - \bar{\delta}_{[k]\setminus\{s\}}\big) + \bar{\delta}_{[k]\setminus\{s\}}.$$2 rather than on an average task (Mestoukirdi et al., 2021, Mestoukirdi et al., 2023).

The theoretical motivation is a weighted empirical-risk formulation

$$\hat{\delta}_s \;=\; \alpha\!\left(k\,\delta_s - (k-1)\,\bar{\delta}_{[k]\setminus\{s\}}\right) + (1-\alpha)\,\bar{\delta}_{[k]\setminus\{s\}} \;=\; \alpha k\big(\delta_s - \bar{\delta}_{[k]\setminus\{s\}}\big) + \bar{\delta}_{[k]\setminus\{s\}}.$$3

with upper bounds that decompose excess risk into an estimation term and a bias term driven by discrepancy distance or Jensen–Shannon divergence. Operationally, the server estimates similarity from gradients at a shared initialization $$\hat{\delta}_s \;=\; \alpha\!\left(k\,\delta_s - (k-1)\,\bar{\delta}_{[k]\setminus\{s\}}\right) + (1-\alpha)\,\bar{\delta}_{[k]\setminus\{s\}} \;=\; \alpha k\big(\delta_s - \bar{\delta}_{[k]\setminus\{s\}}\big) + \bar{\delta}_{[k]\setminus\{s\}}.$$4. Using

$$\hat{\delta}_s \;=\; \alpha\!\left(k\,\delta_s - (k-1)\,\bar{\delta}_{[k]\setminus\{s\}}\right) + (1-\alpha)\,\bar{\delta}_{[k]\setminus\{s\}} \;=\; \alpha k\big(\delta_s - \bar{\delta}_{[k]\setminus\{s\}}\big) + \bar{\delta}_{[k]\setminus\{s\}}.$$5

and client-side noise estimates $$\hat{\delta}_s \;=\; \alpha\!\left(k\,\delta_s - (k-1)\,\bar{\delta}_{[k]\setminus\{s\}}\right) + (1-\alpha)\,\bar{\delta}_{[k]\setminus\{s\}} \;=\; \alpha k\big(\delta_s - \bar{\delta}_{[k]\setminus\{s\}}\big) + \bar{\delta}_{[k]\setminus\{s\}}.$$6, it forms

$$\hat{\delta}_s \;=\; \alpha\!\left(k\,\delta_s - (k-1)\,\bar{\delta}_{[k]\setminus\{s\}}\right) + (1-\alpha)\,\bar{\delta}_{[k]\setminus\{s\}} \;=\; \alpha k\big(\delta_s - \bar{\delta}_{[k]\setminus\{s\}}\big) + \bar{\delta}_{[k]\setminus\{s\}}.$$7

These weights are computed in a special pre-training round in which clients upload mean gradients and variance estimates; afterward, the server runs personalized aggregation.

Because full personalization replaces one downlink broadcast with $$\hat{\delta}_s \;=\; \alpha\!\left(k\,\delta_s - (k-1)\,\bar{\delta}_{[k]\setminus\{s\}}\right) + (1-\alpha)\,\bar{\delta}_{[k]\setminus\{s\}} \;=\; \alpha k\big(\delta_s - \bar{\delta}_{[k]\setminus\{s\}}\big) + \bar{\delta}_{[k]\setminus\{s\}}.$$8 unicasts, both papers introduce clustering of collaboration vectors to limit the number of personalized streams. K-means is run on $$\hat{\delta}_s \;=\; \alpha\!\left(k\,\delta_s - (k-1)\,\bar{\delta}_{[k]\setminus\{s\}}\right) + (1-\alpha)\,\bar{\delta}_{[k]\setminus\{s\}} \;=\; \alpha k\big(\delta_s - \bar{\delta}_{[k]\setminus\{s\}}\big) + \bar{\delta}_{[k]\setminus\{s\}}.$$9, and the number of streams $\alpha=0$0 is selected by a silhouette-based trade-off. Cluster centroids $\alpha=0$1 then define cluster-personalized models

$\alpha=0$2

The empirical record is consistently framed in terms of heterogeneity-aware personalization. In the 2023 paper, full-personalization SCFL achieves average test accuracies of $\alpha=0$3 on EMNIST label shift, $\alpha=0$4 on EMNIST label+covariate shift, and $\alpha=0$5 on CIFAR-10 concept shift. Clustered SCFL with $\alpha=0$6 gives $\alpha=0$7, $\alpha=0$8, and $\alpha=0$9, respectively, while the silhouette score peaks around $\alpha=\tfrac{1}{k}$0 in the label+covariate and concept-shift settings. Worst-user performance is also improved: for EMNIST label+covariate shift with $\alpha=\tfrac{1}{k}$1, Proposed $\alpha=\tfrac{1}{k}$2 reaches 76.4 versus 70.7 for Ditto and 67.5 for FedAvg; for CIFAR-10 concept shift with $\alpha=\tfrac{1}{k}$3, Proposed $\alpha=\tfrac{1}{k}$4 reaches 48.8 versus 43.2 for Ditto and 19.6 for FedAvg. The 2021 paper reports the same server-side idea as a communication-aware personalization scheme and places it explicitly in relation to clustered FL and client-side personalized aggregation.

In this usage, SCFL is not a threat model but a personalization architecture. The “self-centered” qualifier refers to client-specific aggregation weights, not to strategic deviation. That distinction is essential when reading the acronym across papers.

5. Decentralized and representation-sharing variants

“Proximity-based Self-Federated Learning” presents a fully distributed SCFL-style regime in which there is no central server. Clients discover neighbors through a communication graph $\alpha=\tfrac{1}{k}$5, where $\alpha=\tfrac{1}{k}$6 iff $\alpha=\tfrac{1}{k}$7 and $\alpha=\tfrac{1}{k}$8. Neighbor prioritization combines geographic proximity with a symmetric loss-based dissimilarity

$\alpha=\tfrac{1}{k}$9

where $\alpha=1$0. Federations emerge through a multi-leader process and a gradient field $\alpha=1$1 defined as the minimum path sum of dissimilarities to leader $\alpha=1$2; client $\alpha=1$3 joins federation $\alpha=1$4 if $\alpha=1$5 and $\alpha=1$6 is minimal among leaders. Intra-federation aggregation uses FedAvg at the leader, followed by dissemination back to members (Domini et al., 2024).

The reported system is evaluated on EMNIST letters under synthetic label skew across spatial areas $\alpha=1$7 with $\alpha=1$8. The model is an MLP with 128 hidden units, $\alpha=1$9 global rounds, $\ge 3$0 local epochs per round, batch size 64, Adam learning rate 0.001, weight decay 0.0001, and $\ge 3$1. The paper states that PSFL outperforms centralized FedAvg in test accuracy across all tested settings, with larger gains when $\ge 3$2. Lower $\ge 3$3 yields more federations, better alignment with areas, and more stable training; higher $\ge 3$4 can merge dissimilar areas and introduce instability. Additional runs with a moving node show that leader changes and validation NLL spikes are transient and that the federation count returns to the target value.

“Scalable Collaborative Learning via Representation Sharing” provides another client-centric SCFL instantiation, but now with a minimal relay server rather than parameter aggregation. Each client keeps its own model $\ge 3$5, computes per-label averaged representations

$\ge 3$6

uploads $\ge 3$7 and $\ge 3$8 per-class averaged observations, downloads global prototypes

$\ge 3$9

and optimizes

$$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$00

The paper proves the mutual-information lower bound

$$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$01

which makes the contrastive distillation objective well-posed. Communication scales with $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$02, not with model size $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$03. For LeNet5 with $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$04, $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$05, and $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$06, the uplink and downlink are approximately 6.7 KB per client per round. Empirically, on MNIST after 100 rounds, the method achieves 94.19 for $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$07, 90.63 for $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$08, and 82.07 for $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$09, compared with FL at 92.64, 86.79, and 70.06, respectively (Berdoz et al., 2022).

These two variants share a client-centric orientation but embody different system assumptions. Proximity-based self-federation modifies topology and aggregation scope; representation sharing preserves client autonomy by exchanging compact per-label feature summaries. Both depart from the single global-model premise, but they do so through different objects of collaboration: models in one case, representations in the other.

6. Acronym collisions, adjacent clustered/privacy-preserving uses, and open issues

One paper explicitly states that “SCFL stands for Sample Clustered Federated Learning, not Self-Centered Federated Learning.” In that work, SCFL is a semantic-segmentation framework built on Deep Domain Isolation (DDI), which clusters samples rather than clients. DDI computes per-sample, per-class gradients

$$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$10

fits per-class federated Gaussian mixture models, forms sample similarities via the Bhattacharyya coefficient, performs spectral clustering on the server, and then trains one federated model per discovered domain. At inference, a domain classifier routes each sample to the appropriate model. On TMNIST-Inv, the reported test mIoU values are 0.933 for FedAvg in IID and Full splits, 0.970 for CFL in Full, and 0.970 for SCFL in Full; on Cityscapes+GTA5, SCFL reaches 0.775 in Full and 0.780 in Dirichlet. The Rand Index is reported as approximately 0.99–0.996 after sufficient warm-up, and the domain classifier reaches average F1 $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$11 on TMNIST-Inv and approximately 0.997–0.998 on Cityscapes+GTA5 (Manthe et al., 2024).

A different acronym expansion appears in “Social-Aware Clustered Federated Learning with Customized Privacy Preservation.” Here, SCFL is a trust-aware clustered FL system in which mutually trusted users form social clusters, aggregate raw updates within the cluster, and upload only mixed results to the cloud. For low-trust relations, the privacy budget is customized through

$$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$12

with Gaussian local-DP noise scale

$$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$13

Cluster formation is cast as a federation game with a distributed two-sided matching algorithm and Nash-stable convergence. The experiments use a Facebook ego network with 4039 nodes and approximately 88K edges, MNIST with a 4-layer CNN over 30 global rounds, CIFAR-10 with a 5-layer CNN over 100 global rounds, Dirichlet non-IIDness, $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$14, $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$15, $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$16, and $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$17. The paper reports convergence of the cluster-formation game in approximately 7 iterations for $$\delta_{[k]}^t \;=\; \tfrac{1}{k}\sum_{i\in [k]}\delta_i^t, \quad \mathbf{w}^{t+1} \;=\; \mathbf{w}^t + \delta_{[k]}^t.$$18, communication cost per iteration below 5 KB, and partitions evolving from approximately 40 clusters to approximately 25 clusters with approximately 4 users per cluster (Wang et al., 2022).

These acronym collisions are not merely terminological curiosities. They partition the literature into technically distinct questions: strategic self-benefit, personalized aggregation, decentralized federation formation, sample-level domain specialization, and privacy-aware social clustering. A plausible implication is that any citation to “SCFL” without local definition is underspecified.

Across the genuinely self-centered line, several open issues recur. The selfish-client defense literature identifies adaptive selfishness, collusion among selfish clients, fairness implications, and extension to more complex model families and datasets as future work (Augello et al., 2024). FedThief raises the need for defenses against utility-driven, stealth-aligned adversaries that preserve private benefit under robust aggregation (Zhang et al., 30 Aug 2025). The personalization and decentralization papers expose other trade-offs: downlink overhead for per-user models, sensitivity of similarity estimation to noisy gradients, dynamic task drift when weights are computed once, spectral-clustering scalability for sample-level methods, and privacy leakage from shared gradients, responsibilities, or model parameters (Mestoukirdi et al., 2023, Domini et al., 2024, Manthe et al., 2024). As a result, SCFL is best understood not as a single algorithm but as a research area organized around one recurring principle: federated collaboration is reoriented away from a single universal model and toward client-, attacker-, sample-, or trust-specific structure.