Handover-Timed Anytime Updates for Federated Learning on the Move

Federated learning trains one shared model across many devices without collecting their data. Each device computes a local model update and sends it back to a server, which averages the updates. On phones and vehicles that cross small cells every few seconds, the upload is often cut off by a handover, the procedure that moves a device's radio connection from one base station to the next. The half-sent update is usually wasted or badly delayed. This happens even though the cellular network had announced the handover in advance through its own control messages. Before a handover executes, the device reports signal measurements, the network prepares candidate target cells, and an execution condition starts to be met. That information exists and is free, yet no federated learning scheduler uses it.

We propose to use these handover signals as a per-event deadline, and to make every update "anytime." The update is ordered so that whatever prefix gets through before the cut is itself a useful compressed update. The remainder continues through the next cell, tied to the prefix by a few bytes piggybacked on the handover request that base stations already exchange. The goal is to spend less uplink on federated learning in high-mobility deployments, waste almost none of it, and add no new control-plane messages.

Core Claim

Mobile federated learning systems model mobility coarsely. Mobility-Aware Asynchronous Federated Learning with Dynamic Sparsification (MADS) picks how many gradient entries to send from vehicle speed and an assumed exponential distribution of contact time, and it works with a single edge server and no inter-cell handovers. Mobility-Aware Cluster Federated Learning in Hierarchical Wireless Networks (MACFL) lets a client upload to whichever access point it reaches after training, but it sends full weights and ignores what happens to an upload interrupted mid-transfer. Data-Heterogeneous Hierarchical Federated Learning with Mobility shows that mobility helps convergence by mixing regional data across edge servers. It explicitly assumes handovers cost nothing.

Our claim is that three ingredients, used together, reduce the uplink bytes wasted by interrupted transfers and improve accuracy per unit of uplink over these schemes in dense, irregular cell layouts. The first is the timing of the next handover, which the device can read from its own measurement events and conditional-handover state. The second is an update ordered by how much each component changes the model's outputs. The third is a tiny descriptor, carried inside the existing handover request, that lets the target cell's aggregator finish the partial update.

The claim is deliberately narrow. On regular corridors, such as high-speed rail with evenly spaced cells, speed already predicts handover timing and we expect little gain over MADS. The gain should appear where signal quality is volatile. In the operator traces behind Meta-Learning-Based Handover Management in NextG O-RAN, "for 20% of user movements, the SINR can fluctuate by 100% within a single second," so speed is a poor guide to when a link will end.

Research Program

The first component is a deadline estimate. When a device prepares an upload, it computes $\hat\tau$, the predicted time until handover execution, and sets a byte budget

$$B = \min\left(B_{\max},\; \rho\,\hat\tau\,\hat R\right),$$

where $\hat R$ is the currently achievable uplink rate, $\rho<1$ is a safety margin, and $B_{\max}$ is the full compressed update size. The estimate $\hat\tau$ comes from three sources, in increasing order of confidence. The weakest is whether conditional-handover candidates have been prepared. Stronger is how far the target cell's signal exceeds the serving cell's, relative to the execution threshold. Strongest is whether a measurement-event time-to-trigger countdown is running.

Preparation alone is a weak warning. CONTRA found that standard conditional handover prepares many cells it never uses, and removed 77% of those preparations. We therefore train a small predictor on handover traces that maps these signals to a distribution over $\hat\tau$, and we report how its errors affect training. The predictor has the same shape as the per-device temporal models in Inductive Latent Context Persistence, which reports a 99th-percentile inference latency of 7.7 ms. This makes on-device or near-real-time execution plausible.

The second component is the anytime ordering. We reuse the scoring from Enhancing Communication Compression via Discrepancy-aware Calibration for Federated Learning, which ranks update components by how much they shift model outputs on 32–64 local calibration samples. That paper shows that ranking by magnitude, as Top-$k$ does, fails after several local steps. It uses the ranking for a fixed compression ratio. We use it as a transmission order instead, so that cutting the update at any byte count yields the best update of that size. Error feedback keeps unsent components on the device for the next round, which keeps truncation from biasing training.

The third component is cross-cell continuation. When a handover interrupts an upload, the serving cell's aggregator already holds a prefix. We attach a descriptor to the handover request between base stations containing the round number, client identifier, the rank index reached, and a checksum. The target aggregator then accepts the suffix and either merges it or forwards it to the parent aggregator.

This reuses the piggybacking pattern of Inductive Latent Context Persistence, which fits a 128-byte model state into standard handover request messages. We carry bookkeeping for a federated update rather than a radio-control state, and our target is on the order of 16 bytes. No new control messages are introduced. We will count control-plane messages per handover to verify this. We will also confirm that handover failure and ping-pong rates (fast back-and-forth handovers between two cells) are unchanged, because a scheme that helps learning while degrading mobility management would not be deployable.

We then analyze convergence. Truncation at random deadlines behaves like compression with a random, data-independent ratio. We will extend the error-feedback convergence argument to a truncation level set by an independent stopping process. The hypothesis to test is that the resulting bound depends on the expected transmitted fraction, not its worst case.

Baselines

The naive reference is hierarchical federated averaging in which interrupted uploads are dropped and retried in the next round, with a fixed compression ratio. A second naive variant pauses uploads whenever a handover is predicted. Together they show the cost of doing nothing.

The strongest existing method for the compression question is MADS (Mobility-Aware Asynchronous Federated Learning with Dynamic Sparsification). It reports accuracy 8.76% higher on CIFAR-10 and average displacement error 9.46% lower on Argoverse than FedMobile and AFL-Spar. Its sparsity adapts to speed, so it is the direct competitor to adapting to handover events. We extend it to multiple cells by letting it upload to whichever cell it is attached to. It gets the same model, calibration data, uplink rate model, and per-method hyperparameter search budget. We also run an oracle-speed version, because if even oracle speed matches us, our signals add nothing.

The strongest existing method for roaming is MACFL (Mobility-Aware Cluster Federated Learning in Hierarchical Wireless Networks), which reports 86.01% versus 77.95% for standard hierarchical training at moderate mobility on non-IID MNIST (data split unevenly across devices). Because it sends full weights, we also build a stronger hybrid that compresses MACFL uploads at a fixed ratio with discrepancy-aware ranking and resumes interrupted transfers at the target cell. This hybrid is the baseline we most need to beat.

On CIFAR-10 with ViT-tiny at a 0.01 compression ratio, discrepancy-aware calibration reports 29.62% accuracy versus 21.03% for magnitude-based Top-$k$. This anchors the ordering ablation.

Beating the hybrid means reaching a fixed target accuracy with fewer total uplink bytes and less wall-clock time. The margin must be larger than the spread across seeds, and it must hold in the urban and volatile-channel settings.

The ablations isolate each ingredient. The first uses handover-timed budgets with Top-$k$ ordering, which tests the ordering. The second uses discrepancy ordering with a fixed budget and no deadline, which tests the timing signal. The third keeps the timing signal but drops the continuation descriptor, discarding prefixes, which tests continuation. The fourth replaces the learned $\hat\tau$ predictor with conditional-handover preparation alone, which tests whether the cheap signal is good enough.

Why not just…?

Why not just pause training and uploads around handovers? In dense layouts at high speed, the time spent in each cell shrinks toward the length of the handover procedure, as noted in the IAP-MOTION project. Pausing then removes most of the usable uplink time. MACFL reports that standard hierarchical training fails to converge when devices never stay at the same access point. The pause baseline will show how much time is lost.

Why not just set sparsity from speed, as MADS does? This is a real threat. On evenly spaced corridors, speed and handover timing carry nearly the same information. We will report results separately for regular-corridor and irregular urban topologies. If the speed-driven variant, including the oracle-speed version, comes within seed variance of our method in the urban setting, we will conclude that handover signals are unnecessary and say so.

Why not just compress hard and resume interrupted uploads at the next cell? This is the MACFL-plus-compression hybrid, and it may capture much of the benefit. Two things can separate it from our method, and we will measure each directly: the bytes wasted or retransmitted per interruption, and the time until the round's update is fully accounted for.

The first is magnitude ordering. If the hybrid ranks by magnitude, a truncated prefix is a poor update, as the Top-$k$ result above shows. We will use discrepancy ranking in the hybrid so this difference does not decide the comparison. The second is waste. Without a deadline, the device commits to a fixed ratio and often sends a payload too large for the time left in the cell. Without the descriptor, resuming requires re-establishing the session at the application layer, and TCP measurements on high-speed rail show that handover stalls already cause congestion control to throttle throughput.

Why not just use prompt or adapter tuning, so payloads are too small to matter? Parameter-efficient methods such as AUG-FedPrompt bring payloads down to kilobytes. For those models, our method should matter little, and we expect it to. We include a prompt-tuned ViT arm to find the payload size at which the benefit disappears. Larger embedding tables, such as those in FedE, still load edge uplinks heavily, so the regime we target remains relevant.

Proposed Experiments

For mobility, we reproduce the SUMO Manhattan-grid vehicular setup of Data-Heterogeneous Hierarchical Federated Learning with Mobility and overlay a small-cell grid with a 3GPP-style event and conditional-handover model. Channel volatility is set to match the 10–30 dB swings every 10 slots in the CHOMET simulator. We train the $\hat\tau$ predictor, and check handover realism, on the Vienna 4G/5G drive-test traces used in Inductive Latent Context Persistence. Those traces contain only 31 test handovers, so they serve for validation and the simulator carries the main load. If the European operator traces used by CONTRA can be obtained, they supply the volatile-channel stress test. A regular high-speed-rail corridor at 97 m/s is the setting where we expect no gain.

For learning tasks, we use CIFAR-10 with ViT-tiny, to compare against the 29.62%/21.03% calibration numbers; CIFAR-10 with an edge-level non-IID split, to compare with Mob-HierFAVG's 56.0% versus 40.9%; Argoverse trajectory prediction, to compare with MADS; and MNIST with the MACFL split.

For metrics, we report accuracy against uplink bytes and against wall-clock time, and the fraction of uplink bytes discarded or retransmitted. We also report control-plane messages and handover-request bytes per handover, handover failure and ping-pong rates, and the calibration error of $\hat\tau$.

Two checks decide the outcome. First, in the urban volatile setting, our method must beat the MACFL-plus-discrepancy-compression hybrid and oracle-speed MADS on bytes-to-target-accuracy. The ablations must show that the gain comes from the timing signal, not only from the ordering. Second, when we inject deadline prediction errors of increasing size, training must degrade smoothly toward the fixed-budget ablation and never fall below it. If it can fall below, the anytime ordering is not doing its job, and the design must change before any deployment claim.