Client Selection

Selection based on the real-time hardware state of the client node to prevent training failures and ensure computational integrity.

• Charging & Battery Status: Only select devices that are plugged in and above a minimum battery threshold (e.g., >50%) to prevent mid-computation dropout.
• Idle State Detection: Prioritize clients that are currently idle, not actively used by a human, and connected to an unmetered Wi-Fi network.
• Memory & Compute Headroom: Verify sufficient available RAM and CPU/GPU cycles to execute the local training loop without impacting the device's primary function.

Selection based on the communication capability of the client to minimize round time and meet strict communication budgets.

• Throughput Probing: Conduct a lightweight network speed test before selection to estimate upload/download bandwidth and reject clients below a minimum Mbps threshold.
• Latency Sensitivity: Exclude clients with high round-trip time (RTT) to the aggregation server to prevent them from becoming stragglers that delay the synchronous barrier.
• Off-Peak Scheduling: Select clients based on predicted network congestion patterns, scheduling participation during local off-peak hours to maximize throughput.

Selection based on the characteristics of the local dataset to improve global model convergence and mitigate the impact of non-IID data distributions.

• Label Distribution Matching: Select a cohort of clients whose aggregate label distribution closely mirrors the true global distribution to reduce statistical bias.
• Data Volume Filtering: Exclude clients with an insufficient number of local samples, as their updates may represent noisy, high-variance gradients that destabilize training.
• Novelty & Diversity Scoring: Prioritize clients holding data from underrepresented classes or feature spaces to maximize the information gain per communication round.

Selection based on the identity and behavioral history of the client to defend against Byzantine attacks and model poisoning.

• Device Attestation: Verify the cryptographic integrity of the client's runtime environment (e.g., using Trusted Execution Environments) before accepting its updates.
• Reputation Scoring: Maintain a historical ledger of client contributions. Deprioritize or exclude nodes whose past updates consistently deviated from the global consensus or were flagged as anomalous.
• Geographic & Network Diversity: Enforce selection policies that ensure no single autonomous system (AS) or geographic region dominates a round, preventing localized data poisoning.

Selection governed by regulatory and ethical constraints to ensure equitable model performance across diverse populations.

• Mandated Inclusion Quotas: Enforce minimum participation rates for clients representing protected or minority demographic groups to prevent algorithmic bias in the global model.
• Contribution Caps: Limit the maximum number of rounds a single high-resource client can participate in to prevent its local data distribution from dominating the global model weights.
• Randomized Baseline: Always include a small, randomly selected subset of clients to maintain an unbiased baseline for evaluating the impact of strategic selection policies.

Selection based on the mathematical utility of a client's update to accelerate loss reduction, often using multi-armed bandit or gradient-based heuristics.

• Gradient Norm Estimation: Request lightweight, compressed gradient norms from clients before full selection to identify nodes with the largest potential update magnitudes.
• Loss Reduction Prediction: Select clients whose local data distribution is predicted to yield the steepest reduction in the global validation loss, effectively prioritizing the most informative updates.
• Anti-Staleness Weighting: In asynchronous settings, prioritize clients that have most recently synchronized with the global model to minimize the integration of stale gradients.

A set of strategies designed to prevent slow or unresponsive client nodes from bottlenecking the entire federated training round. In the context of client selection, this involves proactively excluding nodes with poor device availability or high latency.

• Deadline-based aggregation: The server ignores updates arriving after a cutoff time
• Coded computation: Introduces redundancy to reconstruct missing updates
• Asynchronous protocols: The server updates the global model immediately upon receiving any single update

The foundational algorithm where a fraction of clients are selected randomly in each round. Selected clients perform multiple local stochastic gradient descent steps before sending their model delta to the central server for weighted averaging.

• Client selection is typically uniform random sampling
• The number of selected clients is a critical hyperparameter
• Serves as the baseline against which advanced selection strategies are benchmarked

A training paradigm where the central server updates the global model immediately upon receiving an update from any single client, eliminating the synchronization barrier. This fundamentally changes client selection dynamics by removing the need to wait for a full cohort.

• Mitigates the impact of stragglers by design
• Introduces gradient staleness as a key challenge
• Selection can be based on

Managing statistical heterogeneity where local client data distributions are not independent and identically distributed. Client selection strategies must account for this to avoid biasing the global model toward the data distribution of over-represented clients.

• FedProx adds a proximal term to limit local model divergence
• SCAFFOLD uses control variates to correct for client drift
• Selection algorithms may enforce statistical diversity in each round's cohort

A hard constraint on the total number of bits or bytes that can be transmitted per client per round or over the entire training run. Client selection directly impacts budget utilization by choosing nodes that can operate effectively under these constraints.

• Drives the adoption of gradient compression techniques
• Selection may favor clients with smaller model delta sizes
• Used to design and benchmark communication-efficient algorithms