Client Selection

Addresses the challenge of Non-IID data across devices. Selection algorithms prioritize clients whose local data distributions contribute to a more generalized global model. Strategies include:

• Greedy selection based on local loss magnitude
• Clustering clients by data similarity before sampling
• Stratified sampling to ensure minority class representation This prevents the global model from overfitting to the data patterns of a dominant subset of users.

Selects clients based on device capability and state to minimize stragglers—slow devices that delay the entire round. The scheduler evaluates:

• Network bandwidth and connection stability
• Battery level and charging state
• Available compute cycles (CPU/GPU/NPU)
• Idle state detection to avoid interrupting active users This ensures training rounds complete within a predictable time window without degrading the user experience on the device.

Incorporates security checks to exclude malicious or faulty nodes before aggregation. Techniques include:

• Anomaly detection on update norms to flag outliers
• Reputation scoring based on historical contribution quality
• Pre-aggregation norm clipping to limit individual influence
• Integration with Remote Attestation to verify device integrity This pre-filtering is a critical defense layer against Model Poisoning and Backdoor Attacks.

Ensures equitable participation and contribution valuation across the client population. Selection logic balances:

• Representation parity to avoid demographic bias in the model
• Contribution tracking for fair reward distribution in incentive systems
• Rotation policies to prevent over-reliance on specific high-value clients
• Federated Unlearning compatibility by tracking which clients contributed to specific model updates This builds trust and encourages sustained participation in the federated ecosystem.

A cryptographic protocol that allows a server to compute the sum of model updates from multiple clients without inspecting individual contributions. Secure aggregation ensures that even if the server is compromised, the raw gradients of selected clients remain private. This is often implemented using Secret Sharing and Masking techniques to protect against gradient leakage during the aggregation phase.

Non-Independently and Identically Distributed data is a common challenge in federated learning where local datasets have heterogeneous statistical distributions. Client selection strategies must account for statistical heterogeneity to avoid biasing the global model. Selecting a diverse, representative subset of clients mitigates the risk of the model overfitting to a specific data distribution.

An attack where a malicious participant manipulates local model updates to corrupt the global model's performance or introduce a backdoor. Robust client selection mechanisms must incorporate norm clipping and outlier detection to identify and exclude anomalous updates before aggregation. This is distinct from data poisoning, which targets the training dataset itself.

A security process verifying the integrity and trustworthiness of a remote client's software and hardware environment before granting access to data or computation. In client selection, remote attestation ensures that only devices running authentic, unmodified code within a Trusted Execution Environment (TEE) are selected to participate, preventing Sybil attacks and compromised nodes from joining the training round.

The foundational federated learning algorithm that combines locally trained model weights on a central server by computing a weighted average. Client selection directly impacts FedAvg's convergence: selecting too few clients increases variance, while selecting too many increases communication overhead. Adaptive selection strategies dynamically adjust the cohort size based on network conditions and data quality.