DynamicFL: Evolving Federated Learning

• DynamicFL is a federated learning approach that explicitly models temporal variations in communication, client participation, and data distributions to enable adaptive protocols.
• It employs techniques such as LSTM-based bandwidth prediction, dynamic clustering, and asynchronous updates to address nonstationary network and data conditions.
• Empirical studies show improved accuracy and reduced training time, though challenges remain in convergence analysis and model consistency.

Searching arXiv for “DynamicFL” and closely related federated-learning frameworks to ground the article in current papers. DynamicFL denotes a family of federated learning formulations in which the federated process is explicitly treated as time-varying rather than static. In the surveyed literature, the label covers bandwidth-aware client selection and client manipulation under fluctuating wireless links, dynamic communication resource allocation that interpolates between FedSGD and FedAvg, client-driven asynchronous personalization under drifting task mixtures, multi-model asynchronous FL under dynamic data drift, and dynamic clustering for fully heterogeneous client architectures (Chen et al., 2023, Le et al., 2024, Li et al., 2024, Chang et al., 2023, Pala et al., 30 Oct 2025). The common thread is that communication conditions, participation patterns, local objectives, model compatibility, or the target optimum itself are allowed to evolve over time.

1. Scope and nomenclature

The term is not standardized. In some works, “DynamicFL” is the explicit name of a framework; in others, the same conceptual territory is covered under adaptive, asynchronous, decentralized, clustered, vertical, or multi-model federated learning. A concise way to read the literature is to distinguish the primary dynamic axis being modeled.

Representative formulation	Dynamic axis	Core mechanism
DynamicFL (Chen et al., 2023)	Communication dynamics in mobile wireless FL	LSTM bandwidth prediction, long-term greedy client selection, adaptive observation window
DynamicFL (Le et al., 2024)	Communication-resource allocation under statistical heterogeneity	High-frequency and low-frequency groups chosen under communication budgets
CDFL (Li et al., 2024)	Client-triggered asynchronous adaptation to evolving task mixtures	Repository of cluster models and server-side mixture estimation
UnifiedFL (Pala et al., 30 Oct 2025)	Fully heterogeneous architectures with evolving optimization similarity	Model-graphs, shared GNN parameter space, dynamic reclustering

This multiplicity matters because a frequent misconception is to treat DynamicFL as a single benign training protocol. In the broader federated literature, dynamic behavior also includes decentralized average tracking without a parameter server and even dynamic adversarial behavior such as changing backdoor targets over time (Chen et al., 2021, Huang, 2020). This suggests that “DynamicFL” is best read as a research direction centered on explicit temporal variation, not as one universally agreed algorithmic template.

2. Communication and resource dynamics

One major interpretation of DynamicFL is system-centric: the bottleneck is dynamic communication rather than static non-IID data alone. In the bandwidth-aware framework of (Chen et al., 2023), the motivating setting is time-sensitive FL over geo-distributed edge devices with highly dynamic wireless bandwidth. The round time is decomposed as

$$T(C_i,R_i)=\Delta \operatorname{Comp}(C_i,R_i)+\Delta \operatorname{Comm}(C_i,R_i),\qquad \Delta \operatorname{Comm}(C_i,R_i)=\frac{\mathbb{U}(pull)+\mathbb{U}(push)}{b_t},$$

so short-term throughput collapse directly turns selected clients into stragglers. The method therefore freezes client selection within an observation window, predicts future bandwidth with a lightweight three-layer LSTM, modifies utility and duration feedback through reward and penalty factors, and adapts the observation-window length through thresholded round-duration control (Chen et al., 2023). In that formulation, the dynamic variable is not the model architecture or objective, but the communication process that determines time-to-accuracy.

A second, more optimization-centric system formulation allocates different synchronization frequencies to different clients. The communication-resource DynamicFL of (Le et al., 2024) partitions active clients into a high-frequency communication group and a low-frequency communication group, with the high-frequency subset chosen so that its joint label distribution is close to the global label distribution. The selection objective is

$\min D_{\text{KL}}\!\left[p(\bar y_z)\|p(Y)\right]$

subject to client and server communication-budget constraints. This construction is intended to bridge FedSGD and FedAvg: some clients communicate frequently enough to act as a drift-correction set, while others remain communication-efficient (Le et al., 2024). A plausible implication is that dynamic communication heterogeneity can be used as a control variable against statistical heterogeneity, rather than being treated only as a systems nuisance.

A related but domain-specific line appears in FLAD for DDoS detection, where the server has no centralized test set for newly emerging attacks. FLAD collects local validation scores $a^c$, forms the mean

$a^\mu=\frac{1}{|C|}\sum_{c\in C} a^c,$

selects only clients with $a^c\le a^\mu$, and assigns more epochs and minibatch-gradient steps to the harder clients through a scaling factor derived from validation performance (Doriguzzi-Corin et al., 2022). Here, dynamic orchestration means adaptive compute allocation and adaptive stopping under privacy constraints, rather than link prediction.

3. Asynchrony, moving objectives, and client-driven personalization

A second major interpretation of DynamicFL is statistical or task-level nonstationarity. CDFL assumes that client $m$ at global epoch $t$ observes a time-varying mixture

$$P_m^t=\sum_{k=1}^K \alpha_{mk}^t P_k,\qquad \sum_{k=1}^K \alpha_{mk}^t=1,$$

over latent cluster distributions. Instead of waiting for server-scheduled rounds, the client independently decides when to upload a locally adapted model; the server maintains a repository of cluster models, estimates the client’s current mixture weights server-side using proxy datasets, updates only relevant clusters, and returns a single personalized model

$u_m^t=\sum_{k=1}^K \hat{\alpha}_{mk}^t w_k^t$

rather than the full model bank (Li et al., 2024). In this view, DynamicFL is fundamentally client-driven and asynchronous: the key problem is rapid re-personalization under evolving local task mixtures.

DMA-FL pushes this further by treating the optimum itself as time-indexed. For each task $j$, the dynamic global objective at aggregation $\min D_{\text{KL}}\!\left[p(\bar y_z)\|p(Y)\right]$0 is

$\min D_{\text{KL}}\!\left[p(\bar y_z)\|p(Y)\right]$1

with instantaneous minimizer

$\min D_{\text{KL}}\!\left[p(\bar y_z)\|p(Y)\right]$2

The server updates asynchronously whenever it receives one trained local model for the task, and the analysis introduces scheduling tensors $\min D_{\text{KL}}\!\left[p(\bar y_z)\|p(Y)\right]$3 and rectangular indicator functions to link asynchronous communication, local training length, and dynamic data drift within one convergence bound (Chang et al., 2023). This formulation makes DynamicFL a tracking problem over a moving target rather than a static ERM problem with delayed gradients.

Dynamic vertical FL provides a narrower but related instance. DVFL considers a passive party whose vertically partitioned data grow over time and can shift in distribution, then uses knowledge distillation to reduce catastrophic forgetting during incremental updates (Liang et al., 2021). This suggests that DynamicFL is not restricted to horizontal cross-device FL; it also includes vertical settings in which the usable aligned dataset itself is temporally evolving.

4. Heterogeneous architectures, decentralized consensus, and dynamic topology

Another branch of the literature treats the federated object itself as structurally dynamic. UnifiedFL addresses fully heterogeneous clients that may use fundamentally different architectures such as CNNs, U-Nets, and MLPs. Each client architecture is rewritten as a directed model-graph $\min D_{\text{KL}}\!\left[p(\bar y_z)\|p(Y)\right]$4, and all clients exchange a fixed-length shared parameter vector $\min D_{\text{KL}}\!\left[p(\bar y_z)\|p(Y)\right]$5 rather than native weights. The server reclusters clients using Euclidean distances

$\min D_{\text{KL}}\!\left[p(\bar y_z)\|p(Y)\right]$6

and applies a two-tier aggregation scheme: frequent intra-cluster averaging and less frequent global averaging over cluster centers (Pala et al., 30 Oct 2025). The dynamic element is thus both representational and topological: similarity groups are recomputed as optimization trajectories evolve.

In decentralized FL, DACFL interprets the missing server-side aggregation step as a dynamic average consensus problem. With a symmetric doubly stochastic mixing matrix $\min D_{\text{KL}}\!\left[p(\bar y_z)\|p(Y)\right]$7, each client updates its consensus state by

$\min D_{\text{KL}}\!\left[p(\bar y_z)\|p(Y)\right]$8

so that $\min D_{\text{KL}}\!\left[p(\bar y_z)\|p(Y)\right]$9 tracks the moving network-wide average model without a parameter server (Chen et al., 2021). Here “dynamic” refers to online tracking of an evolving average, and empirically also to time-varying communication topologies.

A nearby but nonstandard extension appears in collaborative inference with vertically split data over dynamic device environments. That work defines a time-varying graph

$a^c$0

and studies device faults, communication faults, replicated aggregation, gossip rounds, and simulated failures for robust distributed inference and training (Ganguli et al., 2023). Although it is not classical horizontal FL, it belongs to the same DynamicFL landscape because prediction itself depends on which devices and links remain available at inference time.

5. Adaptive aggregation, privacy, and adversarial uses

Dynamic behavior can also be injected directly into aggregation. DRAG computes a global reference direction $a^c$1, measures the angular divergence of each client update through

$a^c$2

and then “drags” the client update toward the reference direction before averaging: $a^c$3 The method is dynamic because the correction is recomputed per client and per round, and it requires no extra communication overhead relative to local SGD/FedAvg (Zhu et al., 2023).

DYNAFED instead mines dynamic information from the early global trajectory $a^c$4. It synthesizes a compact pseudo-dataset $a^c$5 by trajectory matching,

$a^c$6

and then repeatedly fine-tunes the aggregated global model on this pseudo-data in later rounds (Pi et al., 2022). In this usage, “dynamic” refers to leveraging the temporal evolution of the global model, not to changing radio conditions or asynchronous participation.

DearFSAC treats aggregation as a sequential decision problem. The server compresses uploaded local models through an auto-encoder-based quality evaluation network, forms a state from global and local embeddings, local losses, and the previous action, and uses a soft actor-critic policy to output a round-specific aggregation-weight vector

$a^c$7

The dynamic element is state-dependent aggregation under unreliable or defective local updates rather than fixed FedAvg-style averaging (Huang et al., 2022).

Dynamic parameterization under privacy is represented by DP-DyLoRA. A naive per-client dynamic-rank DyLoRA would break differential privacy because clients would send different-length or different-support updates. The fix is to sample one rank on the server for the entire cohort in each round,

$a^c$8

and let all selected clients update the same truncated LoRA factors in that round before clipped Gaussian aggregation (Xu et al., 2024). This is a distinct DynamicFL theme: dynamic adaptation is retained, but the dynamic choice must be cohort-wide to preserve a valid DP mechanism.

Dynamicity is also present on the adversarial side. The dynamic backdoor attack framework in federated learning models training as episodes in which malicious clients can change trigger and target pairs over time, uses a symbiosis network with loss

$a^c$9

and adopts a Reptile-style meta-learning update so that new backdoor tasks can be adapted quickly from prior malicious experience (Huang, 2020). This is not a benign DynamicFL framework, but it is an important reminder that temporal flexibility in FL can be exploited by attackers as well as defenders.

6. Empirical patterns, limitations, and contested scope

Across the literature, the strongest empirical gains appear when the dynamic factor is genuinely consequential rather than nominal. The bandwidth-aware DynamicFL of (Chen et al., 2023) reports better model accuracy while consuming only $a^\mu=\frac{1}{|C|}\sum_{c\in C} a^c,$0–$a^\mu=\frac{1}{|C|}\sum_{c\in C} a^c,$1 of the wall-clock time of the state-of-the-art baseline. The communication-allocation DynamicFL of (Le et al., 2024) reports up to a $a^\mu=\frac{1}{|C|}\sum_{c\in C} a^c,$2 increase in model accuracy and often approaches DynamicSGD with much lower communication cost. UnifiedFL reports that disabling reclustering reduces average F1 by $a^\mu=\frac{1}{|C|}\sum_{c\in C} a^c,$3–$a^\mu=\frac{1}{|C|}\sum_{c\in C} a^c,$4 percentage points across datasets and reduces Dice by about $a^\mu=\frac{1}{|C|}\sum_{c\in C} a^c,$5 percentage points on Hippocampus, indicating that repeated clustering updates, not just architecture unification, drive part of the gain (Pala et al., 30 Oct 2025). FLAD reaches average F1 score $a^\mu=\frac{1}{|C|}\sum_{c\in C} a^c,$6 in $a^\mu=\frac{1}{|C|}\sum_{c\in C} a^c,$7 seconds in its hardest 13-client scenario, compared with markedly slower FedAvg and FLDDoS baselines (Doriguzzi-Corin et al., 2022). DP-DyLoRA reports that with 1 million clients and $a^\mu=\frac{1}{|C|}\sum_{c\in C} a^c,$8, accuracy degradation can be reduced to less than $a^\mu=\frac{1}{|C|}\sum_{c\in C} a^c,$9 and WER increase to less than $a^c\le a^\mu$0 (Xu et al., 2024).

The limitations are equally recurrent. Some methods are system-effective but theoretically light: the bandwidth-aware DynamicFL explicitly provides no formal convergence analysis and contains malformed equations in the utility formulation (Chen et al., 2023). Some provide theory only in highly simplified settings: the communication-resource DynamicFL proves equivalence to FedSGD only in a three-client quadratic model (Le et al., 2024). Some have modeling inconsistencies: UnifiedFL alternates between full $a^c\le a^\mu$1-vectors, topology-plus-parameter descriptors, and compact gradient-moment descriptors when describing reclustering inputs (Pala et al., 30 Oct 2025). Some retain restrictive assumptions: DACFL’s convergence theory assumes i.i.d. data across users (Chen et al., 2021), CDFL assumes a known number of latent clusters and proxy datasets on the server (Li et al., 2024), and DP-DyLoRA remains centralized with secure aggregation and does not analyze client dropout (Xu et al., 2024). Dynamic multi-model wireless formulations such as DMA-FL and DCLM are analytically rich, but they rely on complex mixed-integer non-convex optimization and are evaluated through simulation rather than a deployed wireless testbed (Chang et al., 2023, Abdisarabshali et al., 2024).

The resulting picture is that DynamicFL is best understood as an umbrella for federated systems that explicitly model temporal variation in communication, data, task mixture, topology, aggregation, or parameterization. In some formulations the central challenge is straggler avoidance; in others it is moving-optimum tracking, dynamic clustering, decentralized average tracking, privacy-compatible dynamic rank selection, or robustness to changing attacks. This suggests that the most stable encyclopedia-level definition is not algorithm-specific: DynamicFL is federated learning in which the governing assumptions are dynamic by design, and the algorithmic machinery is built to exploit or control that nonstationarity rather than abstracting it away.