FedSkipTwin: Digital-Twin-Guided Client Skipping for Communication-Efficient Federated Learning

Communication overhead remains a primary bottleneck in federated learning (FL), particularly for applications involving mobile and IoT devices with constrained bandwidth. This work introduces FedSkipTwin, a novel client-skipping algorithm driven by lightweight, server-side digital twins. Each twin, implemented as a simple LSTM, observes a client’s historical sequence of gradient norms to forecast both the magnitude and the epistemic uncertainty of its next update. The server leverages these predictions, requesting communication only when either value exceeds a predefined threshold; otherwise, it instructs the client to skip the round, thereby saving bandwidth. Experiments are conducted on the UCI-HAR and MNIST datasets with 10 clients under a non-IID data distribution. The results demonstrate that FedSkipTwin reduces total communication by 12-15.5% across 20 rounds while simultaneously improving final model accuracy by up to 0.5 percentage points compared to the standard FedAvg algorithm. These findings establish that prediction-guided skipping is a practical and effective strategy for resource-aware FL in bandwidth-constrained edge environments.

💡 Research Summary

FedSkipTwin addresses the communication bottleneck in federated learning (FL) by introducing a server‑side digital twin for each client. Each twin is a lightweight LSTM that learns the temporal pattern of the client’s gradient L2 norm over previous rounds. During every round, the server queries each twin to obtain a prediction of the next gradient magnitude together with an epistemic uncertainty estimate, which is derived via Monte‑Carlo dropout (multiple stochastic forward passes). A dual‑threshold rule is then applied: a client is instructed to skip the round only if both the predicted magnitude is below a magnitude threshold (τ_mag) and the predicted uncertainty is below an uncertainty threshold (τ_unc). This conservative strategy ensures that only updates that are confidently small are omitted, preserving convergence while reducing bandwidth usage.

The algorithm proceeds much like FedAvg. The server broadcasts the current global model, selects clients based on twin predictions, and only those instructed to participate perform E local epochs and send their model delta back. The actual gradient norm of each participating client is fed back to its twin for online retraining, allowing the twin to improve its forecasts as training progresses. Early rounds exhibit low skip rates because twins lack sufficient history; as the global model converges and updates naturally shrink, skip rates increase, yielding dynamic adaptation.

Experiments were conducted on two standard FL benchmarks—UCI‑HAR (human activity recognition) and MNIST—using 10 simulated clients with a non‑IID Dirichlet (α = 0.5) data split. Settings included 20 communication rounds, 3 local epochs per round, batch size 32, and thresholds τ_mag = τ_unc = 0.001 (tuned via grid search). Results show that FedSkipTwin reduces total transmitted data by 15.5 % on UCI‑HAR and 12.0 % on MNIST, while achieving slightly higher final test accuracies (0.9291 vs. 0.9243 on UCI‑HAR, and 0.9669 vs. 0.9656 on MNIST). The convergence curves of FedSkipTwin closely track those of FedAvg, indicating that the skipping mechanism does not destabilize training. Average skip rates across all rounds were 14.8 % (UCI‑HAR) and 11.4 % (MNIST), with the rate rising in later rounds as twins become more confident.

Key contributions are: (1) the novel use of digital twins to predict client‑side update significance, (2) a dual‑threshold, uncertainty‑aware skip rule that conservatively filters low‑impact communications, (3) an implementation on the Flower FL framework demonstrating practical bandwidth savings and modest accuracy gains, and (4) a discussion of how the approach can be combined with existing compression or quantization techniques.

Limitations include the additional computational load on the server for maintaining and updating N LSTM twins, which may become significant in very large‑scale deployments, and the reliance on gradient norm alone as a proxy for update importance, ignoring directionality or layer‑wise contributions. Future work is suggested in three directions: (i) replacing the LSTM with more expressive sequence models (e.g., Transformers) to capture longer‑range dependencies, (ii) enriching the skip decision with multiple metrics such as local loss reduction or parameter drift, (iii) integrating differential privacy mechanisms to prevent twins from inadvertently leaking client‑specific information, and (iv) exploring hybrid schemes that combine twin‑guided skipping with gradient sparsification or quantization for maximal communication efficiency.

Overall, FedSkipTwin demonstrates that predictive, server‑side intelligence can effectively reduce FL communication overhead without sacrificing model performance, offering a promising avenue for resource‑constrained edge scenarios.