Client Selection

The most fundamental policy that filters clients based on their current device state and readiness to participate.

• Charging Status: Selects only devices connected to power to avoid battery drain during compute-intensive training.
• Idle State: Requires the device to be unused and on an unmetered Wi-Fi connection.
• Example: Google's Gboard federated learning only triggers training when the phone is charging, idle, and connected to Wi-Fi.

Policies that evaluate the statistical utility of a client's local dataset before granting participation to prevent model degradation.

• Sample Size Filtering: Rejects clients with too few local examples, which would produce noisy, high-variance updates.
• Label Distribution Check: Ensures the client's local label distribution isn't pathological or adversarial.
• Gradient Norm Screening: Computes a local gradient norm and compares it to the cohort median to detect anomalous data.

Selection logic that accounts for the heterogeneous hardware capabilities of edge devices to minimize straggler effects.

• Compute Tiering: Groups devices by chipset capability and sets per-tier local epoch targets.
• Deadline-Based Selection: Estimates per-client training time and selects only those that can complete within a fixed round deadline.
• Memory Profiling: Excludes devices with insufficient RAM to load the current model architecture.

Advanced policies that intentionally select a representative cohort to ensure the global model doesn't overfit to a dominant subpopulation.

• Clustered Sampling: Groups clients by data distribution similarity and draws proportional samples from each cluster.
• Gradient Diversity Maximization: Selects a subset whose local gradients are maximally orthogonal, ensuring each contributes unique information.
• Importance Weighting: Assigns selection probability inversely proportional to how over-represented a client's data distribution is in the current cohort.

Economic and trust-based mechanisms that govern participation in cross-silo and permissionless federated learning networks.

• Reputation Scoring: Maintains a historical ledger of each client's contribution quality and penalizes those that submit corrupted or low-value updates.
• Staking Requirements: Requires clients to lock a security deposit that is slashed if their update is detected as malicious by a Byzantine-resilient aggregation rule.
• Shapley Value Selection: Prioritizes clients whose historical marginal contribution to model accuracy is highest, as computed by federated data valuation.

Policies that incorporate real-time wireless channel state to optimize communication efficiency in over-the-air federated learning.

• Channel Quality Indicator (CQI) Thresholding: Selects only clients whose current signal-to-noise ratio exceeds a minimum threshold to prevent transmission errors.
• Bandwidth-Adaptive Selection: Adjusts the number of selected clients inversely with available bandwidth to prevent network congestion.
• Over-the-Air Computation Alignment: Selects clients whose channel coefficients naturally align for analog aggregation, maximizing the signal-to-noise ratio of the aggregated update.

The primary challenge driving intelligent client selection. In federated learning, local datasets are statistically heterogeneous—they are not independently and identically distributed. A client may hold data from a single class or a skewed feature distribution. Selecting clients without accounting for this heterogeneity causes client drift, where local updates pull the global model in conflicting directions. Advanced selection algorithms prioritize cohorts whose combined data approximates the global distribution to ensure stable convergence.

Stragglers are slow or unresponsive edge devices that delay synchronous federated rounds. A single device with poor connectivity or limited compute can stall the entire aggregation step. Client selection strategies address this by:

• Deadline-based exclusion: dropping clients that fail to report within a time window
• Resource-aware scheduling: profiling device capabilities (battery, bandwidth, CPU) and selecting only those that can complete training on time
• Asynchronous fallback: incorporating late updates with staleness weighting rather than blocking the round

Client selection must operate under the assumption that some participants may be adversarial or faulty. A Byzantine client can upload arbitrary, malicious updates designed to corrupt the global model. Resilient selection and aggregation employ:

• Robust aggregation rules: coordinate-wise median or trimmed mean instead of simple averaging
• Reputation scoring: tracking historical update quality and deprioritizing anomalous clients
• Statistical outlier detection: comparing each update against the distribution of peer updates before inclusion This ensures the global model converges correctly even when a fraction of selected clients behave arbitrarily.

A targeted attack where a malicious client crafts an update to sabotage the global model or implant a backdoor. Unlike Byzantine faults, poisoning is intentional and often designed to evade detection. In a backdoor attack, the model performs normally on clean inputs but misclassifies inputs containing a secret trigger pattern. Client selection defends against this through:

• Update clipping: bounding the L2-norm of each update to limit influence
• Differential privacy auditing: detecting updates that disproportionately affect specific outputs
• Secure aggregation: preventing the server from inspecting individual updates, forcing attackers to poison the aggregate blindly

The deployment setting that makes client selection both essential and complex. Involves millions of mobile or IoT devices with:

• Intermittent connectivity and unpredictable availability
• Severely limited compute and battery budgets
• Massively non-IID local data (e.g., personal photos, typing patterns) Selection in this regime must be opportunistic: devices volunteer when idle, plugged in, and on unmetered Wi-Fi. The server cannot query all devices; it sets eligibility criteria and accepts updates from those that self-select, creating a fundamentally different scheduling problem from the cross-silo setting.

The game-theoretic foundation for incentive-aware client selection. Not all clients contribute equally—some hold rare, high-quality data that significantly improves model performance, while others contribute redundant or noisy samples. Data valuation quantifies each client's marginal contribution using concepts like the Shapley value. This enables:

• Priority selection: favoring high-value clients when bandwidth is constrained
• Fair compensation: rewarding participants proportionally to their data's utility
• Strategic curation: assembling training cohorts that maximize information gain per round