FedSkipTwin: Digital Twin for Efficient FL

• FedSkipTwin is a federated learning algorithm that leverages lightweight server-side LSTM-based digital twins to predict client update magnitudes and uncertainty.
• It employs a dual-threshold skipping mechanism to selectively request updates, reducing communication by up to 15.5% in bandwidth-constrained, heterogeneous data environments.
• The approach maintains or slightly improves model accuracy compared to FedAvg by filtering out redundant updates, making it practical for mobile, edge, and IoT deployments.

FedSkipTwin is a communication-efficient federated learning algorithm that uses server-side digital twins of clients to decide, round by round, which clients can be safely skipped without significantly affecting learning. It is formulated for a standard cross-device federated learning setting with one central server and $N$ clients, each holding a private local dataset $D_i$, and it targets bandwidth-constrained mobile, edge, and IoT deployments with statistically heterogeneous, non-IID data (Commey et al., 18 Jul 2025). The method augments FedAvg with a prediction-guided skipping mechanism: for each client, the server maintains a lightweight LSTM-based twin that observes a history of update norms, forecasts the next update magnitude and an epistemic uncertainty estimate, and requests communication only when either quantity exceeds a predefined threshold.

1. Problem formulation and motivation

FedSkipTwin is introduced in the context of federated learning where communication overhead remains a primary bottleneck, particularly when full model parameters must be sent in both uplink and downlink at every round. The underlying optimization problem is the standard global objective

$$\min_{\theta} F(\theta) = \sum_{i=1}^{N} \frac{|D_i|}{|D|} F_i(\theta), \quad |D| = \sum_{i=1}^{N} |D_i|,$$

where $F_i(\theta)$ denotes the local loss on client $i$.

The paper considers a standard FedAvg workflow: the server broadcasts the global model $\theta_t$, clients perform local training, and the server aggregates returned updates. In this setting, communication rather than local computation is treated as the dominant systems bottleneck. Prior communication-efficient federated learning methods are described as primarily reducing message size through mechanisms such as sparsification and quantization, while still assuming that all selected clients communicate every round.

FedSkipTwin instead studies “client skipping,” meaning that in a given round some clients are not asked to perform local training and do not upload updates. They simply skip that round, saving both computation and communication. The central question is whether the server can predict, before a round begins, which client updates are likely to be small or redundant and can therefore be omitted with limited effect on optimization. The paper answers this question by modeling each client’s temporal update behavior with a lightweight digital twin.

The experimental setting is explicitly non-IID: data are partitioned among clients using a Dirichlet distribution with concentration parameter $\alpha = 0.5$, producing skewed label distributions. This heterogeneity matters because a naive skipping rule could suppress updates that are rare but informative.

2. Digital twins and the FedSkipTwin protocol

In FedSkipTwin, a digital twin is a server-side virtual model associated with a federated client. For each client $i$, the server maintains a small LSTM network $\text{Twin}_i$. The twin receives as input a sequence of past gradient or update norms for that client and produces two outputs: a prediction of the next update’s $\ell_2$ norm and an estimate of epistemic uncertainty. Because the twins live only on the server, clients incur no additional computation or memory cost beyond standard FedAvg (Commey et al., 18 Jul 2025).

The core per-round quantities are defined through the client update

$D_i$0

and its norm

$D_i$1

For a participating client, the server computes this norm after receiving the update and appends it to the twin’s historical sequence. The twin is then trained online as a time-series predictor.

The overall protocol proceeds as follows. The server initializes the global model $D_i$2 and one digital twin for each client. At each communication round $D_i$3, the server broadcasts $D_i$4 to all clients, or to a pre-selected pool if FedSkipTwin is combined with another client-selection method. The server then queries every twin for a predicted update magnitude and uncertainty estimate. Clients whose predicted updates are sufficiently large, or whose predicted uncertainty is sufficiently high, are asked to train locally and upload updates. The remaining clients skip the round. The server aggregates only the updates from the participating set $D_i$5, and for each participating client its observed update norm is fed back into the corresponding twin for training.

This design makes FedSkipTwin “FedAvg plus a skipping logic.” If all clients communicate, the method reduces exactly to FedAvg. The skipping mechanism is therefore orthogonal to the underlying local optimizer and aggregation rule.

3. Prediction targets, uncertainty estimation, and decision rule

The quantity predicted by each digital twin is the magnitude of the next client update, represented by the $D_i$6 norm of the parameter difference: $D_i$7 The paper uses this norm as a proxy for update significance. Its conceptual argument is that small norms often correspond to redundant or noisy updates, especially later in training, whereas larger norms are more likely to induce meaningful movement in the global parameters.

Epistemic uncertainty is estimated with Monte Carlo dropout. During twin inference, dropout remains enabled and the model performs multiple stochastic forward passes. The resulting distribution of predicted magnitudes is used to form a point prediction and an uncertainty estimate, with the variance of the predictions serving as the uncertainty proxy. The paper states that this is achieved by using Monte Carlo dropout during the twin’s inference step, where multiple stochastic forward passes yield a distribution of predictions whose variance reflects model uncertainty.

The skipping policy is governed by a dual-threshold rule. Let $D_i$8 denote the magnitude threshold and $D_i$9 the uncertainty threshold. Client $$\min_{\theta} F(\theta) = \sum_{i=1}^{N} \frac{|D_i|}{|D|} F_i(\theta), \quad |D| = \sum_{i=1}^{N} |D_i|,$$0 is skipped at round $$\min_{\theta} F(\theta) = \sum_{i=1}^{N} \frac{|D_i|}{|D|} F_i(\theta), \quad |D| = \sum_{i=1}^{N} |D_i|,$$1 only when both its predicted magnitude and its uncertainty are below threshold: $$\min_{\theta} F(\theta) = \sum_{i=1}^{N} \frac{|D_i|}{|D|} F_i(\theta), \quad |D| = \sum_{i=1}^{N} |D_i|,$$2 Equivalently, the server requests an update whenever

$$\min_{\theta} F(\theta) = \sum_{i=1}^{N} \frac{|D_i|}{|D|} F_i(\theta), \quad |D| = \sum_{i=1}^{N} |D_i|,$$3

This rule is explicitly conservative. A client is skipped only when the twin predicts a small update and is simultaneously confident in that prediction. If the predicted magnitude is small but uncertainty is high, the client still communicates. The thresholds are static within a given experiment and are selected by grid search; in the reported experiments, both are set to $$\min_{\theta} F(\theta) = \sum_{i=1}^{N} \frac{|D_i|}{|D|} F_i(\theta), \quad |D| = \sum_{i=1}^{N} |D_i|,$$4: $$\min_{\theta} F(\theta) = \sum_{i=1}^{N} \frac{|D_i|}{|D|} F_i(\theta), \quad |D| = \sum_{i=1}^{N} |D_i|,$$5

A common misunderstanding would be to interpret FedSkipTwin as a purely magnitude-based pruning scheme. The method is not described that way. Its defining mechanism is uncertainty-aware skipping, intended to avoid aggressive omissions when twin predictions are unreliable.

4. Aggregation, systems properties, and practical overhead

For clients that are not skipped, local training is standard FedAvg: the client receives the current global model, runs $$\min_{\theta} F(\theta) = \sum_{i=1}^{N} \frac{|D_i|}{|D|} F_i(\theta), \quad |D| = \sum_{i=1}^{N} |D_i|,$$6 local epochs on its private data $$\min_{\theta} F(\theta) = \sum_{i=1}^{N} \frac{|D_i|}{|D|} F_i(\theta), \quad |D| = \sum_{i=1}^{N} |D_i|,$$7, and returns either its local model $$\min_{\theta} F(\theta) = \sum_{i=1}^{N} \frac{|D_i|}{|D|} F_i(\theta), \quad |D| = \sum_{i=1}^{N} |D_i|,$$8 or the corresponding update $$\min_{\theta} F(\theta) = \sum_{i=1}^{N} \frac{|D_i|}{|D|} F_i(\theta), \quad |D| = \sum_{i=1}^{N} |D_i|,$$9. The server then aggregates only across the participating set $F_i(\theta)$0: $F_i(\theta)$1 This is the standard FedAvg weighting restricted to the clients that did not skip.

The digital twins are implemented as simple LSTM-based time-series models. The paper does not specify hidden size, number of layers, or the exact training loss, but it emphasizes that the twins are simple and lightweight. Their role is deliberately narrow: scalar norm forecasting plus uncertainty estimation. A plausible implication is that the modeling burden is intentionally kept small to preserve scalability at the server.

The server-side overhead is described as negligible. The paper states that each client has an associated twin implemented as a simple LSTM, that LSTMs for scalar time-series prediction are small, and that MC-dropout inference is trivial compared with the main federated model training. On the client side, there is no additional computation or memory beyond standard FedAvg: no extra model, no local prediction mechanism, and no client-side bookkeeping.

The method is also presented as composable with other federated learning control layers. The authors note that FedSkipTwin can be used on top of client selection or other schemes: a system can first choose a candidate client pool based on availability or resources and then apply digital-twin-guided skipping within that pool.

The paper does not provide formal convergence theorems. Its analysis is empirical and qualitative. It argues that gradient norms exhibit temporal structure that small LSTMs can capture, and that the conservative skip rule is designed to preserve convergence by avoiding omission of high-impact or highly uncertain updates.

5. Experimental setup and empirical results

FedSkipTwin is evaluated on two datasets with 10 clients under non-IID partitions generated by a Dirichlet distribution with $F_i(\theta)$2 (Commey et al., 18 Jul 2025). The datasets are UCI-HAR, with 10,299 samples and 6 activity classes, and MNIST, with 70,000 handwritten digit images and 10 classes. The UCI-HAR task uses an MLP with a 128-unit ReLU dense layer, a 64-unit ReLU dense layer, and a 6-unit Softmax output layer. The MNIST task uses a CNN with a 16-filter $F_i(\theta)$3 Conv2D layer, $F_i(\theta)$4 max-pooling, a 32-filter $F_i(\theta)$5 Conv2D layer, another $F_i(\theta)$6 max-pooling layer, flattening, and a 10-unit Softmax dense layer.

Training is conducted for 20 communication rounds with $F_i(\theta)$7 local epochs per round and batch size 32. The implementation uses Python, PyTorch, and the Flower framework. The reported hardware is a Ryzen 7 7700 CPU, 32 GB RAM, and an NVIDIA RTX 3060 Ti GPU.

The main results after 20 rounds are summarized below.

Dataset	Accuracy: FedAvg $F_i(\theta)$8 FedSkipTwin	Communication: FedAvg $F_i(\theta)$9 FedSkipTwin
UCI-HAR	$i$0	$i$1
MNIST	$i$2	$i$3

On UCI-HAR, communication is reduced by $i$4, while final accuracy improves from $i$5 to $i$6, a gain of $i$7 or $i$8 percentage points. On MNIST, communication is reduced by $i$9, while final accuracy improves from $\theta_t$0 to $\theta_t$1. Across the two datasets, the paper reports a total communication reduction of $\theta_t$2–$\theta_t$3 over 20 rounds together with equal or slightly better final accuracy.

The convergence curves show that FedSkipTwin closely tracks FedAvg on both datasets, with no noticeable slowdown despite using fewer updates. In later rounds, the FedSkipTwin curve lies slightly above FedAvg, consistent with the final accuracies. An efficiency plot of accuracy versus total communication places FedSkipTwin in a more favorable region, combining higher accuracy with lower communication volume.

The per-round skip rate is dynamic rather than fixed. Early rounds exhibit low skip rates because the twins are poorly trained and predictions are uncertain. As training progresses, gradient norms tend to decrease, the twins improve, and the average skip rate rises. The reported average skip rates across all rounds are $\theta_t$4 of client-rounds skipped for UCI-HAR and $\theta_t$5 for MNIST.

The paper’s discussion also implies an ablation-style conclusion, although not in a formal ablation table: uncertainty-aware skipping is important because skipping solely on predicted magnitude would risk missing rare but important updates, especially under heterogeneous data distributions.

6. Limitations, interpretation, and relation to broader “twin” methods in federated learning

Several limitations are explicitly acknowledged. First, threshold tuning currently relies on grid search, and the selected thresholds remain fixed throughout an experiment. The paper suggests adaptive thresholding as future work. Second, the digital twins suffer from a cold-start problem: in early rounds, little history is available, so predictions and uncertainty estimates can be poor. Suggested remedies include better priors, transfer learning for twin initialization, and shared or tiered twins for similar clients. Third, scaling to very large federated systems may become challenging because the server must maintain one per-client model, even if each model is small. Finally, the method relies on the predictability of gradient norms; in settings with highly erratic client behavior or abrupt changes in local data, predictive quality may degrade (Commey et al., 18 Jul 2025).

The paper also offers an interpretation for the slight accuracy improvements over FedAvg. Later in training, many client updates are small and may primarily represent noise, especially in non-IID settings. Skipping low-magnitude, confidently predicted updates can therefore reduce the influence of noisy updates on the global model and may act as a form of implicit regularization. This interpretation is presented as intuition rather than theorem.

Within the broader federated learning literature, “twin” terminology is used in materially different ways. In "Application-Aware Twin-in-the-Loop Planning for Federated Split Learning over Wireless Edge Networks" (Ding et al., 28 Apr 2026), the digital twin is a cross-domain planner composed of network, training, and task sub-twins for federated split learning under wireless resource constraints. In "Overcoming Catastrophic Forgetting in Federated Class-Incremental Learning via Federated Global Twin Generator" (Nguyen et al., 2024), the “twin” consists of a pair of server-side generators that synthesize images and features for class-incremental replay. FedSkipTwin uses neither formulation. Its twin is specifically a per-client, server-side LSTM that predicts update magnitude and uncertainty to control round-level participation.

Taken together, these distinctions place FedSkipTwin in a specific sublineage of digital-twin-assisted federated learning: server-centric, client-lightweight, and oriented toward communication scheduling rather than task-level planning or generative replay. Its declared contribution is to show that prediction-guided communication scheduling, driven by lightweight per-client digital twins, can reduce total communication by $\theta_t$6–$\theta_t$7 over 20 rounds while maintaining or slightly improving final model accuracy in non-IID, bandwidth-constrained edge environments.