Client Selection

The statistical rule determining a client's chance of being picked. Uniform random selection gives every device an equal probability, providing an unbiased sample but ignoring resource heterogeneity. Proportional selection weights probability by a utility metric like available compute cycles, battery level, or local dataset size. A well-designed distribution balances statistical representativeness against the risk of over-representing high-resource clients, which can skew the global model toward their data distributions.

A pre-selection gate that excludes clients incapable of completing the training task within a deadline. Filtering criteria include:

• Minimum battery threshold: Rejects devices below 20% charge to prevent mid-round dropout.
• Network bandwidth floor: Ensures clients can upload model updates within the straggler timeout.
• Compute capability check: Verifies sufficient CPU/GPU to finish local epochs on time. This deterministic filtering reduces straggler-induced delays but risks introducing systemic bias by excluding an entire class of low-resource devices.

The degree to which the selected client subset mirrors the true global data distribution. In Non-IID environments, uniform random sampling can produce a cohort whose aggregated data diverges significantly from the population. Advanced selection strategies use clustering-based sampling or stratified selection to ensure each statistical stratum of the client population is proportionally represented. Poor representativeness causes client drift, where local updates pull the global model in conflicting directions and slow convergence.

A scheduling mechanism that maintains a temporal fairness ledger for each client. Algorithms like Round-Robin with Staleness Bounds ensure no client is starved of participation for too long, while capping how frequently a single client contributes to prevent overfitting. Key metrics tracked:

• Time since last selection: Prioritizes clients with the longest gap.
• Update staleness: Deprioritizes clients whose last contribution is too old to be relevant. This prevents the global model from oscillating between the data distributions of a small, frequently selected subset.

Selection strategies designed to maximize the reduction in global loss per communication round. Gradient-based selection prioritizes clients whose local gradients have the largest magnitude or are most orthogonal to the current global update direction, maximizing information gain. Loss-based selection targets clients with the highest local training loss, focusing computation where the model performs worst. These methods accelerate convergence but require clients to share metadata like loss values or gradient norms, introducing a minor privacy-utility trade-off.

The property of a selection algorithm to maintain model integrity when a fraction of clients are adversarial or faulty. Robust aggregation-compatible selection avoids deterministic patterns that an attacker could exploit to control the cohort. Techniques include:

• Randomized subset sampling: Prevents adversaries from predicting and coordinating their inclusion.
• Reputation-weighted selection: Deprioritizes clients whose historical updates were statistical outliers. This is critical in cross-device FL where compromised IoT sensors could inject poisoned updates.