Spending the Energy Hole: Federated Learning on Multi-Hop Sensor Networks

Battery-powered sensor networks forward everything toward a single collection point, the sink. The nodes next to the sink relay traffic for everyone, so they die first and the network goes dark while nodes at the edge still hold 50–90% of their battery. Networking researchers have learned to make use of this stranded energy. Edge nodes wake up more often, transmit farther, or add redundant packets, and none of this shortens the network's life, because that life is set by the overloaded relays anyway. Federated learning, where devices train a shared model on local data and send only model updates, is now being pushed onto exactly these networks. Yet it still treats battery as a limit to stay under, never as a surplus to spend.

This note proposes using the stranded edge energy for learning work. Edge nodes would run more local training, send their updates over long-range links that skip the overloaded relays, and pre-combine their neighbors' updates. The proposal also argues that this only helps if the server's averaging is corrected for the imbalance it creates. Nodes that did more work, or whose updates reached the sink by a different route, would otherwise pull the shared model toward the data at the network's edge. The contribution is the pairing of an energy-surplus allocation rule with an aggregation rule that removes the bias this allocation introduces. It is tested on networks where relays die and the routing tree reshapes itself over time.

Core Claim

Hold network lifetime fixed, defined as the time until the most loaded relay near the sink dies. Under that constraint, a federated learning protocol that spends each node's projected surplus energy on extra local optimization and on bypass links should reach higher accuracy than an energy-uniform protocol. The claim depends on a second condition: the gain should appear only when aggregation corrects for unequal local work and for subtrees whose updates arrive late or not at all. Under spatially correlated data, where edge sensors see different conditions than central ones, the sharper prediction is this. Plain sample-weighted averaging combined with energy-driven extra work should hurt accuracy on the central region. Directly up-weighting energy-rich nodes, the obvious design, should hurt it further.

That second half is what separates this from the obvious idea at the intersection. The obvious idea is "energy-rich clients contribute more, so weight them more." The research file supports energy as a reason to assign work. It gives no support for energy as evidence that an update is good. The proposal therefore keeps energy out of the aggregation weights and uses it only to decide who does what work.

Relation to Prior Work

Three lines of work motivate the energy side, and all of them operate below the learning layer. ODC-SFNS raises duty cycles of energy-rich nodes in sparse regions, cutting delay by 20.52–79.96% with energy utilization above 90%. RDML lets edge nodes switch to a longer transmission radius that bypasses intermediate relays, raising lifetime by up to 84.40% and energy utilization from 23.33% to 88.00%. URL adds coding redundancy only far from the sink. We reuse RDML's long-radius bypass and ODC-SFNS's idea of a sparse-region surplus as concrete spending actions. These papers move sensor readings, not model updates, and model updates differ in two ways. They are large, and they can be averaged in the network, so a relay can combine its children's models into one before forwarding. Whether the energy hole even survives under in-network averaging is an open empirical question that we answer first.

Two papers connect energy to federated learning, and both constrain our design. REWAFL chooses participants by residual energy and link rate and reaches 0% dropout. It assumes every device talks directly to the server, and it uses energy to include or exclude clients, not to assign extra work. CEHFed uses UAVs as mobile intermediate aggregators and treats their battery only as a disconnection risk to manage by repositioning. We differ from both by modeling multi-hop relay load and by deliberately spending surplus.

Hat-DFed is the closest work on the topology side. It uses bandits to choose a sparse peer-to-peer topology under energy budgets and weights neighbor models by validation loss. It found that about 30% connectivity was best. Its energy model is symmetric across nodes, and it runs on fixed ground servers with no sink-centered bottleneck.

On aggregation, FedVal scores updates on a small server validation set. FedISM weights clients by loss flatness and warns that loss-based reweighting can drive the model into sharp minima. FedISM also reports that its sharpness-aware local training doubles local compute. That cost makes it a natural candidate for surplus-only execution, which we test as a secondary spending option.

Research Program

The first step is measurement. We build a multi-hop simulator covering both RDML's k-ary routing trees and ODC-SFNS's circular 500 m deployment. Model-update traffic is costed in bytes per hop, and computation in joules per local step. On this simulator we compare two forwarding schemes for federated rounds. In end-to-end forwarding, each update travels to the sink separately. In in-network averaging, each relay combines its subtree before forwarding.

For each scheme we report the per-hop energy profile and the fraction of initial energy left unused when the first relay dies. If in-network averaging flattens the hole so that less than about 20% of energy is stranded, the premise weakens. In that case the program narrows to end-to-end forwarding settings and to mobile-sink settings, where the hole is known to persist.

The second step is the surplus estimator. Each node $k$ tracks its residual energy $E_k(t)$ and its baseline drain rate $\rho_k$. Using routing-tree information its parent relays down, it also tracks the projected remaining lifetime $T(t)$ of the most loaded relay on its path. Its spendable surplus is

$$S_k(t) = E_k(t) - \rho_k\, T(t) - \epsilon_k,$$

where $\epsilon_k$ is a safety reserve. In words, $S_k(t)$ is the energy the node would still hold when the network dies anyway.

Surplus is spent through three actions. Extra local steps are the cheapest. A long-radius uplink that skips one or more relays saves energy for the relays it bypasses. Acting as a sub-aggregator for siblings means fewer, combined models enter the bottleneck. A node with negative surplus runs a minimum number of local steps and forwards only. Because relays die and the tree reroutes, $T(t)$ and therefore who counts as peripheral change during training. Every allocation is recomputed each round from current routing state.

The third step is aggregation, which is what makes asymmetric work safe. Each partial aggregate carries three running sums: weighted per-step updates, sample mass, and sample-weighted step counts. Local updates are normalized by step count before combining. For the set $A$ of nodes whose updates actually reached the sink, the server applies

$$\Delta = \frac{\sum_{k\in A} n_k \tau_k}{\sum_{k\in A} n_k}\cdot\frac{\sum_{k\in A} n_k\, \Delta_k/\tau_k}{\sum_{k\in A} n_k}.$$

Here $n_k$ is node $k$'s sample count, $\tau_k$ its number of local steps, and $\Delta_k$ its model change. Normalizing removes the pull toward nodes that took more steps. Because $A$ contains only arrivals, a subtree cut off by a dead relay drops out cleanly instead of silently re-weighting the remaining nodes.

Lost subtrees still bias the objective over many rounds. To counter this, the server keeps each subtree's last arrived contribution and blends it in with a staleness discount. This plays the role of RTEG's storage links, which buffer data until a route reappears, but at the level of model updates. As an optional quality layer, nodes with large surplus may run FedISM-style sharpness-aware local training, and their flatness score may enter aggregation as FedISM prescribes. This tests whether surplus can buy the robustness that FedISM says is otherwise too expensive at the edge.

Baselines

The naive reference is plain sample-weighted federated averaging over the same multi-hop tree. Every node runs the same fixed number of local steps and the same radius, with the step count tuned so that the first relay death falls at the same time as for our method. Matching the lifetime makes the comparison one of accuracy at equal lifetime, not accuracy at different lifetimes.

The strongest selection-based competitor is REWAFL, ported to the multi-hop tree. Its energy utility is computed from end-to-end per-round cost, including relay forwarding. REWAFL reports 0% dropout, 37.9–51.2% faster time-to-accuracy than Oort, and 18.2–54.9% energy savings on its four-dataset suite (MNIST, CIFAR-10, Shakespeare, HAR). Those numbers come from a star topology. There is no reported result on Lacuna for REWAFL or any federated method on a multi-hop sensor tree with a sink bottleneck, so its standing in our setting has to be measured. It is the strongest competitor because it already turns battery state into a learning decision. If simply choosing well matches our results, then spending surplus adds nothing.

The strongest topology competitor is Hat-DFed, run as decentralized learning over the same physical radio graph, with its bandit topology construction and validation-loss aggregation. It reports an average 1.9% accuracy gain and 36.7% energy reduction over D-PSGD and DAC on CIFAR-10. It is a real threat because removing the sink removes the hole.

All methods share the same data partitions, radio and energy model, model architecture, and number of hyperparameter configurations for tuning, chosen by validation accuracy at the lifetime point.

Ablations isolate the key ingredient. "Surplus without correction" keeps our allocation but uses plain sample-weighted averaging. "Energy up-weighting" weights each update by $n_k$ times residual energy. "Correction without surplus" uses uniform work with our aggregation. "Bypass only" and "extra steps only" separate the two spending actions.

We count the proposal as beating the field under two conditions. First, it must exceed both REWAFL-ported and Hat-DFed in accuracy at equal lifetime by more than seed-to-seed spread across five seeds. Second, "surplus without correction" must lose accuracy on the central region relative to the full method.

Why not just…?

Why not just select energy-rich clients, as REWAFL does? Selection decides who trains but leaves the relays carrying whatever the selected nodes send. A selected edge node's update still crosses the overloaded relays, so selection alone cannot move load away from them. Selection also excludes central nodes, whose data is spatially distinct. FedVal's finding that uniform treatment leaves FEMNIST minority accuracy at 32% versus 53% with adaptive scoring suggests that dropping distinct groups costs accuracy. That evidence comes from a different mechanism, so we measure the risk directly with worst-region accuracy. The REWAFL-ported baseline settles this question. If it matches us, the spending actions are not pulling their weight.

Why not just raise local steps for everyone? At fixed lifetime, the relays cannot afford more steps, so uniform increases shorten lifetime or must be capped at the relays' budget. The "correction without surplus" and naive baselines, tuned to equal lifetime, show the best accuracy uniform allocation can reach. If extra steps at the edge add nothing beyond that tuned optimum, the claim fails.

Why not just use step-normalized averaging and skip the energy logic? Normalization fixes bias but generates no extra useful work, so on its own it should match the naive baseline when work is uniform. The ablation runs exactly this configuration. If it matches the full method, the energy allocation is unnecessary, and we would report that.

Why not just go decentralized and avoid the sink? This is the most serious alternative. Peer-to-peer averaging has no single bottleneck, and Hat-DFed shows sparse topologies are efficient. Many sensor deployments, however, need a sink for actuation or uplink, and multi-hop gossip over a radio graph has its own relay costs. Running Hat-DFed on the identical radio graph decides it. If it wins, the honest conclusion is that sink-centered federated learning should be abandoned in these networks, and the surplus estimator would be retargeted to gossip.

Proposed Experiments

Data come from HAR, the sensor-like dataset REWAFL uses, from FEMNIST and Shakespeare in the LEAF suite, and from CIFAR-10. Partitions are spatially correlated. Users or Dirichlet label shards are assigned by ring distance from the sink, so edge and central data differ systematically, and an IID control is run alongside.

Networks are RDML-style k-ary trees and ODC-SFNS-style 500 m circular deployments with 100–1,000 nodes. Dynamic topology enters in three ways. Relays die and the tree reroutes. Random churn removes nodes independently. A mobile sink or UAV collector follows a CEHFed-like trajectory, so bottleneck position moves over time.

Metrics are global and worst-ring accuracy at first relay death, accuracy per joule, the fraction of initial energy used at network death (directly comparable to RDML's and ODC-SFNS's energy-utilization figures), and rounds completed.

Three checks are decisive. The phase-one measurement must show a stranded-energy fraction large enough to matter. The full method must beat the REWAFL-ported and Hat-DFed baselines at equal lifetime. "Surplus without correction" and "energy up-weighting" must measurably degrade central-ring accuracy under spatially correlated data and not under IID data. That pattern is the signature that the correction is doing the work the proposal claims for it.