[Paper] Feature-Based Semantics-Aware Scheduling for Energy-Harvesting Federated Learning

Source: arXiv - 2512.01983v1

Overview

The paper introduces a semantics‑aware client scheduling technique for federated learning (FL) on devices that harvest energy from the environment. By estimating how much a local model update will actually improve the global model, the method trims unnecessary computation, which is often the dominant energy drain on edge devices. The authors achieve this with a lightweight proxy for the Version Age of Information (VAoI), making the approach practical for real‑world, energy‑constrained FL deployments.

Key Contributions

• VAoI‑based scheduling: Leverages a semantics‑aware freshness metric (VAoI) to prioritize updates that are both timely and informative.
• Feature‑based proxy: Replaces the costly full‑parameter statistical distance with a single‑pass intermediate‑layer feature extraction, cutting computation by orders of magnitude.
• Energy‑aware client selection: Integrates harvested‑energy constraints directly into the scheduling decision, preventing wasteful local training when the device cannot afford it.
• Extensive evaluation: Demonstrates superior model accuracy and up to ~30 % energy savings under highly non‑IID data and severe energy scarcity, compared to existing EH‑FL baselines.

Methodology

1. Problem formulation – Each edge device harvests stochastic energy and must decide whether to (a) train a local DNN update, (b) transmit it, or (c) stay idle. The goal is to maximize global model performance while respecting the device’s energy budget.
2. Version Age of Information (VAoI) – VAoI measures how “stale” a model version is and how much its parameters differ from the current global model (i.e., the semantic relevance of the update). Traditional VAoI requires computing a distance over the entire parameter space, which is prohibitive on low‑power hardware.
3. Feature‑based proxy – The authors extract activations from an intermediate layer of the DNN after a single forward pass of a small validation batch. The Euclidean distance between these feature vectors serves as a surrogate for the full‑parameter distance, capturing redundancy with far less compute.
4. Scheduling algorithm – At each communication round, the server collects (i) each client’s harvested energy state and (ii) the proxy VAoI estimate. It then solves a lightweight knapsack‑like selection problem, picking the subset of clients whose expected contribution per joule is highest.
5. Training loop – Selected clients perform local SGD for a fixed number of epochs, then upload their updates. The server aggregates them (FedAvg) and repeats.

Results & Findings

Interpretation: By discarding updates that are semantically redundant, the system converges faster and uses significantly less energy. The feature‑based proxy incurs negligible extra latency, confirming its suitability for on‑device execution.

Practical Implications

• Edge AI deployments (e.g., smart cameras, wearables) can now run FL without draining their tiny batteries, because they avoid wasteful local training cycles.
• Network operators gain finer control over uplink traffic: only the most “valuable” updates are transmitted, reducing congestion in massive‑IoT scenarios.
• Framework integration – The proxy can be wrapped as a lightweight PyTorch/TensorFlow hook, making it easy to plug into existing FL libraries (Flower, PySyft, FedML).
• Energy‑harvesting hardware designers can use the scheduling logic to dimension solar/thermal harvesters more accurately, knowing the algorithm will self‑throttle when energy is scarce.

Limitations & Future Work

• The proxy relies on a fixed intermediate layer; its effectiveness may vary across model architectures (e.g., transformers vs. CNNs).
• Experiments focus on image classification benchmarks; broader domains such as NLP or time‑series forecasting remain untested.
• The scheduling problem is solved greedily; exploring more optimal (but still lightweight) combinatorial solvers could further improve performance.
• Extending the approach to asynchronous FL and multi‑task settings is an open research direction.

Authors

• Eunjeong Jeong
• Giovanni Perin
• Howard H. Yang
• Nikolaos Pappas

Paper Information

• arXiv ID: 2512.01983v1
• Categories: cs.LG, cs.DC, cs.IT, cs.NI, eess.SP
• Published: December 1, 2025
• PDF: Download PDF