Feature-Based Semantics-Aware Scheduling for Energy-Harvesting Federated Learning

Abstract

Federated Learning (FL) on energy-harvesting edge devices is constrained by the high computational cost of model training relative to available energy. This paper introduces a lightweight, semantics-aware scheduling framework that utilizes a feature-based proxy to predict the value of local updates before committing to energy-intensive backpropagation. By extracting feature vectors through a single forward pass, the system prioritizes informative updates, reducing energy consumption by up to 37% while maintaining high accuracy in non-IID, energy-constrained environments.

Introduction

In modern Internet of Things (IoT) ecosystems, Federated Learning (FL) enables decentralized model training while preserving data privacy. However, many edge devices rely on intermittent renewable energy (Energy-Harvesting FL, or EHFL). While communication bandwidth is a traditional bottleneck, the computational energy required for training deep neural networks (DNNs) often far exceeds transmission costs.

Existing EHFL strategies frequently employ "blind" selection policies, choosing clients based on random sampling or battery levels. This leads to redundant updates where precious energy is wasted on data that does not significantly improve the global model. The challenge lies in predicting a client's contribution efficiently before the expensive training process begins.

The VAoI-Aided Scheduling Framework

The proposed framework addresses these challenges through a scheduling mechanism based on Version Age of Information (VAoI). VAoI is a semantic metric that tracks both the freshness of a client's information and its deviation from the current global model.

Expand

Source, p. 2

As shown in Figure 1, the scheduling process involves:

1. Global Model Broadcast: The server distributes the current model $\mathbf{w}^{(t)}$ to all clients.
2. Scheduling Decision: Clients use the VAoI-aided scheduling algorithm to decide whether to participate based on battery levels ($E_i^{(s)}$) and data significance.
3. Local Training: If selected (Case 1), the client performs local training to update its model $\mathbf{w}_i^{(t)}$.
4. Local Update Upload: The client uploads its update to the server if energy conditions are met.
5. Server Aggregation: The server aggregates updates to produce $\mathbf{w}^{(t+1)}$.

Feature-Based Proxy for Semantic Awareness

To implement semantic awareness without high computational overhead, the authors developed a Feature-Based Proxy. Instead of comparing high-dimensional parameter vectors, the client performs a single forward pass on a small batch of local data.

The process involves:

• Feature Extraction: Extracting a feature vector from an intermediate hidden layer.
• Statistical Distance: Calculating the distance ($M_i^{(t)}$) between the current feature vector and a historical average from the last training event.
• Decision Logic: If the distance exceeds a threshold, the update is deemed "semantically significant," and the device commits to the full training process.

Experimental Evaluation

The framework was evaluated using the CIFAR-10 dataset across 100 clients. To simulate realistic edge heterogeneity, data was partitioned using a Dirichlet distribution with varying concentration coefficients ($\alpha$). Lower $\alpha$ values represent more extreme data heterogeneity. The energy model assumed training costs 20 units of energy, while transmission costs only 1 unit.

Key Results and Insights

The VAoI-aided scheme demonstrates significant improvements in both energy efficiency and learning performance compared to standard baselines like FedAvg and battery-aware policies like FedBacys.

Expand

Source, p. 5

Energy Efficiency and Performance

The proposed method reduced network-wide energy consumption by up to 37% compared to FedAvg. As illustrated in Figure 5, the VAoI-aided approach (Ours) consistently outperforms baselines, particularly in scenarios with severe data heterogeneity ($\alpha = 0.1$) and low energy availability ($p_{bc} = 0.01$). By spending limited energy on the most impactful data, the system maintains a more accurate and synchronized global model.

Conclusion

This paper presents a shift from "saving bits" to "saving flops" in edge-based Federated Learning. By utilizing a lightweight semantic proxy, the VAoI-aided framework enables efficient energy usage on resource-constrained devices. Future work will explore adaptive layer selection for feature extraction and dynamic thresholding to further optimize performance across diverse neural architectures.

@misc{jeong2025feature,
  author = {Eunjeong Jeong and Giovanni Perin and Howard H. Yang and Nikolaos Pappas},
  title = {{Feature-Based Semantics-Aware Scheduling for Energy-Harvesting Federated Learning}},
  year = {2025},
  eprint = {2512.01983},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.01983},
}