---
title: Performance-Based Client Selection Strategy
url: https://www.emergentmind.com/topics/performance-based-client-selection-strategy-pbcs
type: topic
---

# Performance-Based Client Selection Strategy

Searching arXiv for the cited papers and closely related client-selection work to ground the article in current literature.
Performance-Based Client Selection Strategy (PBCS) denotes a class of federated learning (FL) client-selection methods in which participation is determined by explicit performance signals predictive of training utility, system efficiency, or service-level objective (SLO) compliance rather than by uniform random sampling alone. Across the literature, PBCS encompasses deterministic top-\(K\) selection by utility score, stochastic probability-weighted selection, correlation-aware selection, deadline-aware chance-constrained selection, and self-selection mechanisms driven by local losses or other client-side metrics. In the most explicit formulation, the server evaluates available clients by a utility score reflecting their expected contribution and resource profile, selects the top performers, and aggregates their updates within a standard FL objective \(F(w)=\sum_{i=1}^{N}\pi_i f_i(w)\) [2502.00036]. More broadly, the term covers methods that use local loss, gradient norms, gradient alignment, Shapley-value contribution estimates, or GPU deadline-compliance probabilities as selection signals [2010.01243] [2012.08009] [2111.11204] [2103.13822] [2312.09108] [2403.17833] [2511.08147]. The common rationale is that selective participation can improve convergence, accuracy, latency, communication efficiency, privacy–utility trade-offs, or fairness under data and system heterogeneity.

## 1. Definition and conceptual scope

In FL, a PBCS prioritizes clients according to a measurable signal associated with expected global improvement, round efficiency, or both. Typical signals include local loss, accuracy, gradient magnitude, gradient alignment with a global direction, validation utility, or application-specific proxies such as firing-intensity shifts in spiking federated learning [2103.13822] [2010.01243] [2111.11204] [2403.17833] [2406.12200]. The essential departure from uniform random selection is that selection is biased, intentionally, toward clients judged more useful under the current optimization and system state.

The formulation in "Efficient Client Selection in Federated Learning" [2502.00036] is a representative server-side PBCS. The paper evaluates each available client \(i \in A_t\) in round \(t\) by a utility \(u_i^t\) reflecting expected contribution to model improvement and round efficiency, with performance signals drawn from data quality and computational capacity. A generic expression consistent with that paper is
\[
u_i^t = w_Q \cdot Q_i^t + w_R \cdot R_i^t,
\]
where \(Q_i^t\) is a data-quality metric and \(R_i^t\) is a resource metric, after which the server performs deterministic top-\(K\) selection
\[
S_t = \mathrm{TopK}(\{(i,u_i^t)\}_{i\in A_t}, K_t).
\]
This specific instantiation integrates differential privacy (DP), fault tolerance (FT), and an adaptive number of selected clients \(K_t\) [2502.00036].

Other works broaden the notion of “performance” beyond data quality and capacity. "Client Selection in Federated Learning: Convergence Analysis and Power-of-Choice Selection Strategies" formalizes biased selection toward clients with higher local losses and analyzes how such bias alters convergence speed and solution bias [2010.01243]. "Bandit-based Communication-Efficient Client Selection Strategies for Federated Learning" models client selection as a discounted UCB problem using observed local losses as rewards [2012.08009]. "Client Selection in Federated Learning based on Gradients Importance" instead scores clients by gradient norms \(\|g_i(w_t)\|_2\) [2111.11204]. "GPFL: A Gradient Projection-Based Client Selection Framework for Efficient Federated Learning" uses the projection of a client’s local gradient onto the global descent direction,
\[
c_i^t = \frac{\langle \nabla \ell_i(w_i^t), g^{t-1}\rangle}{\|g^{t-1}\|},
\]
thereby treating descent alignment as the performance criterion [2403.17833].

A plausible implication is that PBCS should be understood less as a single algorithm than as a design principle: client participation is decided by predictive utility estimates under explicit optimization, systems, or fairness objectives. That principle is instantiated differently depending on whether the dominant constraint is heterogeneity, latency, privacy, energy, fairness, or communication budget.

## 2. Core mathematical structure

Most PBCS methods remain embedded in the standard FL optimization objective. In [2502.00036], the global objective is
\[
F(w)=\sum_{i=1}^{N}\pi_i f_i(w),
\]
with \(\pi_i\) commonly taken as \(n_i/\sum_j n_j\), and after selection the server aggregates selected updates as
\[
w^{t+1}=w^t-\eta\sum_{i\in S_t}\alpha_i^t \tilde{g}_i^t.
\]
The default aggregation weights are data-size based,
\[
\alpha_i^t=\frac{n_i}{\sum_{j\in S_t} n_j},
\]
although uniform weighting among selected clients is also consistent with the described algorithmic steps [2502.00036].

Performance-based selection modifies the stochastic sampling process through a selection rule, explicit score, or constrained optimization. In deterministic top-\(K\) PBCS, the server constructs a binary selection mask \(s_i^t\in\{0,1\}\), with \(p_i^t=s_i^t\) for deterministic rounds [2502.00036]. In probability-based variants, the selection probability itself is the key quantity. "ProbSelect: Stochastic Client Selection for GPU-Accelerated Compute Devices in the 3D Continuum" derives a deadline-compliance probability
\[
p_i = \mathbb{P}(T_i + C_i \le t_{\mathrm{SLO}})
\]
from an analytical GPU latency model and a prior over realized FLOP efficiency \(\eta_i \sim \mathcal{N}(\mu_i,\sigma_i^2)\), then either thresholds on \(p_i\) or samples clients proportionally to it [2511.08147]. Its chance-constrained objective is
\[
\max_{S: |S|=K} \sum_{i\in S}\mathbb{P}(T_i + C_i \le t_{\mathrm{SLO}}),
\quad
\text{s.t. }\mathbb{P}(T_i + C_i \le t_{\mathrm{SLO}})\ge \alpha \;\forall i\in S.
\]

Loss-based PBCS can also be written as biased selection over clients with larger \(F_k(w)\). In [2010.01243], the effect of such selection is analyzed through a selection-skew quantity \(\rho(S(\pi,w),w')\), which measures how much the strategy overweights clients with larger local objectives relative to unbiased sampling. That work shows that larger skew can improve the decaying term in the convergence bound while introducing a non-vanishing bias term [2010.01243].

Correlation-aware PBCS, as in FedCor, replaces independent client utility with a posterior expectation over inter-client loss-change correlations. FedCor models
\[
\Delta \ell^t \sim \mathcal{N}(\mu^t,\Sigma^t),
\]
and selects clients to minimize the posterior-weighted expected next-round global loss [2103.13822]. Shapley-based PBCS, by contrast, scores clients through marginal utility contributions \(\phi_i\), using validation-set performance of aggregated subsets as the set function \(v(S)\) [2312.09108].

This diversity of formulations indicates that the central mathematical object in PBCS is not a fixed score formula but a mapping from client-specific observables to expected marginal utility under FL constraints.

## 3. A canonical PBCS workflow in federated learning

The workflow in [2502.00036] provides an explicit round-level template. At round \(t\), the inputs are the global model \(w^t\), available clients \(A_t\), privacy parameters \((C,\sigma,\varepsilon)\), checkpoint interval \(t_c^*\), time and communication budgets \((\tau_t, B_t)\), and client-count bounds \((K_{\min},K_{\max})\). The round then proceeds through performance estimation, adaptive client-count selection, client selection, local training with clipping and DP noise, upload and aggregation, and privacy-accounting update [2502.00036].

The performance-estimation stage computes or collects \(Q_i^t\) and \(R_i^t\) for each available client, then ranks clients by \(u_i^t\). The adaptive-count stage chooses \(K_t\) subject to
\[
\mathbb{E}[T_t(K_t)] \le \tau_t,
\qquad
\mathrm{Comm}_t(K_t)\le B_t.
\]
A practical formalization consistent with the paper’s narrative is to choose \(K_t\) to maximize a round utility under time and communication constraints, and to increase \(K_t\) when validation improvement plateaus but prior round time remains below budget, while decreasing \(K_t\) under stragglers or timeouts [2502.00036].

Local training then applies DP at the client level through gradient clipping and Gaussian noise:
\[
\bar{g}_i^t = g_i^t \cdot \min\{1, C/\|g_i^t\|_2\},
\qquad
\tilde{g}_i^t = \bar{g}_i^t + \mathcal{N}(0,\sigma^2 I).
\]
The server aggregates the received \(\tilde{g}_i^t\) and updates the privacy accountant using the sampling rate \(q_t=K_t/N\) [2502.00036].

The same paper ties PBCS to a simple synchronous round-time model,
\[
T_t \approx \max_{i\in S_t} T_{\mathrm{comp}}(i) + T_{\mathrm{comm}}(K_t) + T_{\mathrm{overhead}}(\mathrm{FT},\mathrm{DP}),
\]
and communication cost
\[
\mathrm{Comm}_t \approx K_t \times (|w| + |update|).
\]
This is significant because the method’s performance-based ranking is not purely statistical: higher-capacity clients reduce \(\max_i T_{\mathrm{comp}}(i)\), while adaptive \(K_t\) constrains communication and straggler exposure [2502.00036].

A plausible implication is that this workflow functions as a baseline architecture for more specialized PBCS variants. Deadline-aware probability selection [2511.08147], bandit-based selection [2012.08009], or gradient-projection selection [2403.17833] can be interpreted as alternative choices for the performance-estimation layer.

## 4. Major variants and representative mechanisms

The literature on PBCS can be organized by the type of performance signal used and by whether selection is deterministic, stochastic, centralized, or decentralized.

| Variant | Core signal | Representative paper |
|---|---|---|
| Utility-based top-\(K\) | Data quality + capacity | [2502.00036] |
| Loss-based biased sampling | Local loss or stale/discounted loss | [2010.01243], [2012.08009] |
| Correlation-aware selection | GP posterior over loss-change correlations | [2103.13822] |
| Gradient-based selection | Gradient norm or gradient projection | [2111.11204], [2403.17833] |
| Contribution-based selection | Shapley-value proxy | [2312.09108] |
| Deadline-aware probabilistic selection | SLO compliance probability | [2511.08147] |
| Self-selection | Locally computed loss thresholding | [2602.06601] |

Loss-based selection has two influential forms. Power-of-Choice samples a candidate pool of size \(d\), queries their local losses \(F_k(w^{(t)})\), and selects the \(m\) highest-loss clients, yielding up to \(3\times\) faster convergence and \(10\%\) higher test accuracy than random selection in the reported experiments [2010.01243]. UCB-CS removes pre-round polling by treating each client as an arm in a discounted UCB process whose reward is average local loss over \(\tau\) local steps, thereby achieving faster convergence and higher fairness with only a scalar loss report per selected client [2012.08009].

Gradient-based PBCS replaces loss with update geometry. Gradient-importance selection ranks clients by \(\|g_i(w_t)\|_2\), and the paper proves a standard nonconvex convergence rate for the single-client case under bounded gradients and a positive-correlation assumption [2111.11204]. GPFL, by contrast, uses global-direction alignment and augments it with a UCB-style exploit–explore term to prevent over-selection of the same clients [2403.17833].

Correlation-aware PBCS emphasizes complementarity rather than individual value. FedCor models client loss-change correlations with a Gaussian Process and selects clients that minimize expected next-round global loss after conditioning on predicted loss reductions of selected clients. The paper reports convergence-rate improvements of \(34\%\)–\(99\%\) on FMNIST and \(26\%\)–\(51\%\) on CIFAR-10 over Power-of-Choice [2103.13822].

Contribution-based methods use set functions. GreedyFed computes approximate Shapley values \(\phi_i\) using server-side validation utility \(v(S)\), then greedily selects high-contributing clients under timing constraints [2312.09108]. FedGCS generalizes contribution-based selection further by learning a continuous representation of client subsets and optimizing a multi-objective comprehensive score involving performance, latency, and energy before generating the final subset [2405.06312]. This suggests a broader trend: PBCS increasingly appears as structured combinatorial optimization rather than merely score ranking.

Probabilistic and decentralized variants extend the concept beyond server-side top-\(K\). ProbSelect analytically computes per-device GPU efficiency thresholds and transforms them into compliance probabilities \(p_i\), selecting clients likely to satisfy a deadline without historical monitoring [2511.08147]. Type-based unsourced FL instead lets clients self-select based on locally computed loss \(\ell_k^{(t)}\) and a broadcast threshold \(\theta^{(t)}\), with participation probability
\[
\Pr\{s_k^{(t)}=1 \mid \ell_k^{(t)}\} = \sigma(a(\ell_k^{(t)}-\theta^{(t)})),
\]
thus avoiding server-side access to client-specific utilities [2602.06601].

## 5. Privacy, fault tolerance, fairness, and systems constraints

A distinctive feature of the PBCS in [2502.00036] is its explicit coupling of selection with DP and FT. DP is applied per client by clipping and Gaussian noise, with experiments varying the privacy budget \(\varepsilon\) and showing that higher \(\varepsilon\) yields higher accuracy on both UNSW-NB15 and ROAD. The same paper notes that the sampling rate \(q_t=K_t/N\) affects privacy accounting, so aggressive client selection can consume privacy budget faster because larger \(K_t\) increases \(q_t\) [2502.00036]. This links PBCS directly to privacy expenditure rather than treating privacy as an orthogonal layer.

The FT mechanism is based on checkpointing with interval \(t_c^*\). Each selected client periodically saves local state and, upon failure, resumes from the latest checkpoint. The server may wait within a timeout or proceed with partial participation, depending on policy. Reported FT overhead is \(5\%\)–\(10\%\) in training time, with a modest \(2\%\)–\(3\%\) decrease in accuracy or AUC under failures [2502.00036]. This is notable because it demonstrates that PBCS can be embedded in a systems-robust FL stack without requiring asynchronous or fully redundant execution.

Fairness is more contentious. Performance-based top-\(K\) risks repeatedly selecting high-capacity or high-quality clients, thereby under-representing slower or statistically atypical clients. The paper on PBCS in [2502.00036] explicitly states that it does not report fairness metrics or constraints and notes that, in practice, one can mitigate bias via exploration terms, caps on per-client selection frequency, or fairness-aware terms in \(u_i^t\). Similar concerns recur throughout the literature. UCB-CS frames exploration as a fairness mechanism and reports higher Jain’s index than several baselines [2012.08009]. LongFed formulates the performance–fairness tension as a Lyapunov-constrained long-term optimization over diversity in gradient space and individual fairness among similar clients [2405.13584]. UNIONFL uses a history-coverage penalty to discourage repeated participation of recently selected clients [2408.13683].

This suggests that fairness in PBCS is not a peripheral issue but a structural consequence of biased participation. The tension is particularly sharp when performance signals correlate with data quality, device speed, or network conditions. HAPS-enabled selection by traffic class and latency-aware clustered FL both prioritize homogeneous or fast clients under wireless constraints, which improves convergence but may over-represent specific client types unless classes are rotated or quotas are imposed [2401.05308] [2108.08768].

## 6. Empirical behavior, trade-offs, and misconceptions

The empirical results in [2502.00036] provide one of the clearest performance summaries for PBCS as an integrated framework. On the network anomaly detection datasets UNSW-NB15 and ROAD, the method improves accuracy by \(7\%\) and reduces training time by \(25\%\) relative to the baselines ACFL and FedL2P [2502.00036]. Under the DP study, accuracy on UNSW-NB15 rises from \(86\%\) at \(\varepsilon=10\) to \(89\%\) at \(\varepsilon=100\), while ROAD rises from \(73\%\) to \(82\%\), demonstrating the expected privacy–utility trade-off [2502.00036]. Under FT, UNSW-NB15 accuracy changes from \(94.8\%\) to \(92.1\%\), AUC from \(0.93\) to \(0.91\), and training time from \(570\)s to \(600\)s; ROAD shows analogous modest degradations [2502.00036].

These numbers are consistent with a broader empirical pattern. Power-of-Choice reports up to \(3\times\) faster convergence and about \(10\%\) higher test accuracy than random selection [2010.01243]. FedCor reports large convergence-rate gains under non-IID partitions [2103.13822]. GreedyFed improves stability and accuracy under timing constraints and heterogeneity [2312.09108]. GPFL reports over \(9\%\) improvement in FEMNIST test accuracy in non-IID settings [2403.17833]. ProbSelect improves SLO compliance by \(13.77\%\) on average and reduces computational waste by \(72.5\%\) for MobileNetV2 relative to a deadline-aware random baseline [2511.08147].

Several misconceptions recur in discussions of PBCS. One is that selecting “better” clients always means selecting higher-loss clients. The literature is more heterogeneous. Power-of-Choice and UCB-CS do prioritize high-loss clients [2010.01243] [2012.08009], but mmFedMC, in a multimodal FL setting, reports better behavior from selecting lower-loss clients per modality because higher-loss selection can oscillate between heterogeneous local minima [2401.16685]. Another misconception is that PBCS is necessarily server-centric. The self-selection strategy in unsourced FL demonstrates the opposite: clients can decide locally using only a threshold broadcast by the server, while the server remains identity-agnostic [2602.06601].

A further misconception is that PBCS is equivalent to system-aware scheduling. In fact, some methods are explicitly optimization-centric and largely ignore system cost, such as FedCor’s GP loss-correlation model [2103.13822], while others are system-centric and prioritize deadline compliance or latency, such as ProbSelect [2511.08147] or clustered wireless CFL [2108.08768]. PBCS is best seen as the umbrella under which these objectives are encoded into client utility.

## 7. Limitations, open questions, and future directions

Several papers note the absence of formal convergence guarantees for their specific PBCS instantiations. The PBCS in [2502.00036] explicitly states that no formal convergence bounds are provided. It suggests, under standard smoothness and bounded-variance assumptions, that prioritizing higher-quality updates could improve constants in typical FL convergence rates while DP noise adds variance terms proportional to \(\sigma^2\), but this remains an interpretive rather than proved result [2502.00036]. Similar gaps exist for GreedyFed [2312.09108], FedGCS [2405.06312], and SFedCA [2406.12200].

Scalability is another open issue. Methods that rely on exact or approximate Shapley values, GP correlation models, or learned generative selection spaces can deliver high-quality selection but may become expensive at large client counts [2103.13822] [2312.09108] [2405.06312]. Probabilistic analytic methods like ProbSelect are attractive partly because they avoid continuous monitoring and historical telemetry in the 3D edge–cloud–space continuum [2511.08147]. This suggests a likely bifurcation between lightweight score-based PBCS for large-scale cross-device FL and richer combinatorial PBCS for cross-silo or moderate-scale settings.

Fairness, incentive compatibility, and adversarial robustness remain active themes. SBRO-FL combines Shapley values, bids, and prospect-theoretic reputation in a budget-constrained integer program, showing that performance-based selection can be fused with economic and trust signals [2505.21219]. LongFed and UNIONFL show that fairness can be built into the objective via Lyapunov queues or historical coverage penalties rather than added as an afterthought [2405.13584] [2408.13683]. A plausible implication is that future PBCS work will increasingly treat selection as a multi-objective control problem spanning optimization, systems, privacy, and social constraints.

Finally, PBCS is no longer confined to standard FL. Spiking FL uses firing-intensity changes as a native client credit signal [2406.12200]. Multimodal FL jointly optimizes modality and client selection, with client selection based on per-modality local loss [2401.16685]. Wireless unsourced FL uses client self-selection under privacy-preserving identity-free aggregation [2602.06601]. These extensions indicate that PBCS has become a general design language for selective participation in distributed learning, rather than a narrow heuristic confined to classical FedAvg.

In that sense, PBCS marks a shift from participation as random sampling to participation as controlled, model-aware, and system-aware decision-making. The literature consistently shows that this shift can improve efficiency and robustness, but it also exposes new tensions—especially around fairness, privacy accounting, and selection bias—that increasingly define the frontier of federated learning research.

Source: https://www.emergentmind.com/topics/performance-based-client-selection-strategy-pbcs