Standard Federated Learning (FL) relies on a comfortable assumption: every edge device has a reasonable, uniform chance of participating in the training process. In the real world, this assumption quickly falls apart. Hardware limitations, erratic network connectivity, varying battery lives, and localized user activity patterns mean that client participation is highly non-uniform. Some devices participate frequently, some drop out intermittently, and others are systematically excluded, never connecting to the central server at all.
When client participation is incomplete or biased, the fundamental machine learning guarantees we take for granted begin to break down.
The Theoretical Wall: Why Incomplete Participation Breaks Learnability
To understand why incomplete participation is catastrophic, we have to look at the mismatch it creates between the training data distribution and the true population distribution.
If a specific subpopulation of clients—say, users with older phone models or limited connectivity—is permanently blocked from participating, their data remains completely invisible to the central server. In a series of foundational analyses, Haibo Yang et al. proved in papers like With a Little Help from My Friend: Server-Aided Federated Learning with Partial Client Participation and Understanding Server-Assisted Federated Learning in the Presence of Incomplete Client Participation that standard FL frameworks fail to be Probably Approximately Correct (PAC) learnable under worst-case incomplete participation.
In plain terms, this means that even if the participating clients provide an infinite amount of data and train for an infinite number of iterations, the global model's generalization error will never converge to zero. The permanent exclusion of a client segment creates an irreducible distribution shift.
Even when clients are not permanently excluded but merely participate intermittently, a severe generalization gap emerges. As Zheshun Wu et al. demonstrate in Advocating for the Silent: Enhancing Federated Generalization for Nonparticipating Clients, standard aggregation algorithms implicitly optimize for the "loudest" or most frequently available devices, leaving the passive or silent participants with a model that generalizes poorly to their local data.
Restoring Learnability via Server Assistance
If the server cannot learn a representative model because entire portions of the population distribution are hidden, how do we bypass this theoretical impossibility? One powerful solution is to change the architecture of the federated system itself by introducing a small amount of server-side data.
Standard FL:
[Hidden Clients] ✖───> (No Data) ───> [Server Model] <─── (Biased Data) <─── [Active Clients]
│
(Irreducible Bias)
Server-Aided FL:
[Hidden Clients] ✖───> (No Data) ───> [Server Model] <─── (Biased Data) <─── [Active Clients]
▲
│ (Anchors & Corrects Bias)
[Auxiliary Dataset]
In On the Efficacy of Server-Aided Federated Learning against Partial Client Participation, Haibo Yang et al. show that equipping the central server with a tiny auxiliary dataset completely revives PAC-learnability guarantees. This framework, known as Server-Aided Federated Learning (SA-FL), uses the server's clean, representative dataset as an anchor.
To make this practical, the authors introduced SAFARI, an algorithm analyzed across multiple iterations of their work, including Understanding of Server-Assisted Federated Learning with Incomplete Client Participation. SAFARI uses the server-side dataset to compute auxiliary gradients that correct the directional bias of the updates coming from the skewed, active client pool. By doing so, it matches the communication efficiency of ideal FL while robustly correcting for the systematic exclusion of edge devices.
Balancing the Scales for Intermittent Clients
When clients are not completely hidden but instead suffer from unpredictable, non-uniform availability, the goal shifts from introducing external data to intelligently rebalancing the data the server does receive.
Dynamic History Tracking
If the server blindly aggregates updates using standard Federated Averaging (FedAvg), the global model will naturally skew toward clients with stable, high-speed connections. To counter this without incurring massive memory overhead, Shiqiang Wang et al. proposed FedAU (Federated Averaging with Adaptive Weighting) in A Lightweight Method for Tackling Unknown Participation Statistics in Federated Averaging. FedAU tracks the historical participation frequency of each client on the fly. When an infrequently available client finally manages to connect, the algorithm dynamically scales up its aggregation weight, ensuring its rare contribution carries enough mathematical weight to offset the influence of the frequent participants.
Risk-Aware Objectives
Another approach is to alter the optimization goal itself to protect vulnerable users. In Federated Learning under Restricted user Availability, Periklis Theodoropoulos et al. introduced a risk-aware objective function. Instead of optimizing for average performance across all received updates, this objective explicitly accounts for the variance in user availability, forcing the training process to prioritize performance stability on less accessible users and stabilizing an otherwise volatile training trajectory.
Embracing Stale Updates
Finally, instead of discarding updates from clients that fail to complete their training within a strict synchronous time window, we can choose to value delayed information. In FedStale: leveraging Stale Updates in Federated Learning, Angelo Rodio et al. developed a strategy to safely integrate stale model updates into the global aggregation. By combining fresh updates with delayed updates from slower or temporarily disconnected devices, FedStale reduces the variance and systemic bias introduced by rigid synchronization cutoffs, ensuring that slow devices are not effectively silenced.
Key Foundational Work
With a Little Help from My Friend: Server-Aided Federated Learning with Partial Client Participation (2022) — contributed 4 concept elements to this direction.
On the Efficacy of Server-Aided Federated Learning against Partial Client Participation (2023) — contributed 3 concept elements to this direction.