Traditional Federated Learning (FL) operates under a "closed-world" assumption: a fixed set of clients is available, and the server simply picks a subset for each round. In reality, edge devices are volatile. They drop out due to poor signals, join the network halfway through training, or only participate when plugged into power.
This research direction focuses on moving FL from static, idealized environments to dynamic systems where the participant pool is in constant flux. Researchers are developing mechanisms to ensure that model convergence, privacy, and fairness remain intact even when the "who" of the training process is a moving target.
The Challenge of Client Churn
In a standard FL round, the server sends a model to clients, waits for updates, and aggregates them. If a client drops out mid-update, their contribution is lost. If this happens frequently, the global model becomes biased toward "reliable" clients (e.g., those with better hardware or stable Wi-Fi), leading to a collapse in fairness and accuracy for the rest of the population.
Qinbin Li and colleagues in their survey on federated learning systems highlight that most frameworks struggle with this instability. When nodes join or leave during active rounds, the system must decide how to integrate new knowledge without overwriting what was learned from now-absent participants. This is particularly difficult in "online" settings, where Yun Xin and Jianfeng Lu note in DaringFed that the server must make recruitment decisions sequentially without knowing the quality of future participants.
Predicting Reliability through Trust and Scheduling
One way to handle volatility is to stop treating all clients as equally available. Instead of random sampling, researchers are building "trust-aware" systems that predict which clients are likely to finish a round.
Gangqiang Hu and Jianmin Han introduced TRAIL, a framework that uses an Adaptive Hidden Semi-Markov Model to predict a client's "reliability duration." By scheduling associations based on these predictions, the system avoids "flaky" clients that would otherwise stall convergence. Similarly, Minxue Tang and Yu Wang proposed FedCor, which models the correlations between clients to select groups that maximize global improvement, even when connectivity is intermittent.
Decoupling Participation from Aggregation
A significant bottleneck in dynamic FL is the cryptographic "handshake." Many secure aggregation protocols require a fixed set of users to agree on keys at the start of a round; if one leaves, the whole round might fail.
To solve this, Nazatul H. Sultan and Yan Bo developed RLSA-PFL, which decouples the masking process from global pairwise agreements. This allows participants to join or leave at any round without breaking the security of the aggregation. In the realm of privacy, Andreas Athanasiou and Catuscia Palamidessi explored metric-privacy, which scales noise based on the distance between updates rather than a fixed global constant, making privacy guarantees more robust to a shifting participant pool.
Maintaining Knowledge Across Churn
When a client leaves the network, their unique data distribution often leaves with them. This is a major hurdle for "incremental" learning, where the model is supposed to learn new tasks over time.
Tae Jin Park and Dimitrios Dimitriadis addressed this with FIL-VER, which uses "Variational Embeddings" to perform server-side rehearsal. This allows the global model to "remember" the data characteristics of departed clients. On the optimization side, Michael Crawshaw and Mingrui Liu developed Amplified SCAFFOLD, an algorithm specifically designed for periodic and stochastic availability. It uses long-range control variates to ensure that even if a client only appears once every dozen rounds, their update doesn't cause the global model to drift.
Efficiency in Unstable Networks
Finally, the sheer size of modern models (like LLMs) makes dynamic participation even harder. If a device has a narrow window of connectivity, it cannot afford to download and upload gigabytes of parameters.
Frameworks like FEDPRUNER by Yebo Wu and Li Li tackle this by pruning redundant layers, allowing edge devices to participate in fine-tuning with 75% less memory. Similarly, Boyu Li proposed Adaptive Federated LoRA, which adjusts the "sampling probability" and "sketching ratio" of updates based on real-time wireless conditions, ensuring that even brief windows of connectivity are used effectively.