Keeping Federated Clients From Quitting: Training Under Risk-Driven Dropout

Federated learning trains a shared model without collecting anyone's data. Each phone, hospital or bank trains on its own records and sends back only a model update. The whole approach depends on people and institutions agreeing to keep taking part. The federated learning literature treats a client that stops participating as bad luck: a dead battery, a lost connection, or a scheduled lockout. Technology-adoption research gives a different account. People abandon high-stakes systems when their perceived risk rises, and that perception is driven by things they can see, such as crashes, lag and unclear privacy promises. If federated clients behave the same way, the server itself is partly causing the dropout, through how slow, glitchy or leaky each round feels.

This note proposes to study federated training when dropout depends on perceived risk, and to test one specific claim. When clients leave permanently because of risk signals the server controls, correcting for missing updates after the fact is the wrong remedy. The server should instead manage each client's risk exposure during training so that the clients whose data is hardest to replace stay in the federation. We will build a calibrated simulator of risk-driven dropout, a server-side retention scheduler, and a decisive comparison that separates the effect of why clients leave from how many leave.

Core Claim

Existing dropout-robust methods assume that the chance a client participates does not depend on what training does to that client. Debiasing Federated Learning with Correlated Client Participation models participation as a Markov chain set by lockout rules and rescales each client's learning rate by its inverse participation frequency. Friends to Help: Saving Federated Learning from Client Dropout replaces a missing update with the update of a similar active client. Stabilizing and Improving Federated Learning with Non-IID Data and Client Dropout (ReBaFL) uses class priors and server-side feature prototypes so the model does not lose track of classes whose clients have vanished. All three treat dropout as coming from outside the training process, and the research file confirms that no FL paper models dropout as a decision driven by client risk.

The adoption literature suggests three reasons that assumption fails in practice. Users read visible instability as evidence of hidden insecurity: in Is Metaverse Technology Ready to Welcome Online Banking Users?, 13 of 19 empirical studies name perceived risk as the main barrier, and technical fragility triggers fear of financial loss. Perceived risk turns into "technology fear" that suppresses use (Long et al.). Formal safeguards and clear information lower perceived risk (Chen Wen-pei and Feng Yan; Cui et al.).

If these findings transfer, three consequences follow. Dropout becomes permanent churn rather than intermittent absence. It becomes correlated with the data, because clients holding sensitive or rare records are plausibly the most risk-averse. And it becomes a feedback loop, because slower or leakier rounds cause more departures.

Under permanent churn, reweighting cannot recover a client who never returns. Inverse-frequency weights for that client either blow up or become undefined. Friend substitution also breaks down when a whole cluster of similar clients leaves together, which is exactly the pattern that correlated, risk-driven departures produce. Our claim is that a server which models each client's departure risk and spends its limited levers on the clients whose data is hardest to replace will beat post-hoc correction on accuracy over the full population. The levers are lighter local workloads, fewer or more reliable rounds, and targeted privacy protection. Full-population accuracy means accuracy measured on data from all clients, including those who left.

Research Program

The first component is a risk-driven dropout model. For client $k$ in round $t$, the probability of leaving permanently is

$$h_{k,t} = \sigma\big(\alpha_k + \beta_L L_{k,t} + \beta_F F_{k,t} + \beta_P P_{k,t} - \beta_A A_k\big),$$

where $\sigma$ is the logistic function. The terms are defined as follows. $\alpha_k$ is the client's baseline risk-aversion, drawn so that it can correlate with the client's label mix. $L_{k,t}$ is the wall-clock latency the client has experienced recently. $F_{k,t}$ counts recent failed or aborted rounds. $P_{k,t}$ is the privacy exposure disclosed to or inferable by the client. $A_k$ indicates whether a structural assurance, such as an audit statement or a named accountable aggregator, is present, as proposed conceptually in Aligning Federated Learning with Existing Trust Structures in Health Care Systems.

For $P_{k,t}$ we use a cheap stand-in for the invertibility risk score of From Risk to Resilience. That paper measures how easily a client's data can be reconstructed from its update using singular values of the model's input Jacobian. It reports that reconstructions become unrecognizable below a score of 0.15 and high-fidelity above 0.45. We will not claim that users perceive that score directly. It serves as the objective quantity that a disclosure could communicate.

The coefficients $\beta$ are unknown. The file contains no measured link between federated system events and real departure, and the adoption literature measures stated intention rather than logged behavior. That weakness is flagged by A Critical Review of Theory of Planned Behavior in Knowledge Payment. The literature is also internally tense. Long et al. find that ease of use has no significant effect on adoption, while Ma et al. find that instability matters a great deal. Latency could therefore matter a lot or not at all.

We handle this uncertainty in two ways. First, every simulation result is reported across a grid of $\beta$ values, never at a single chosen point. Second, we run a small calibration study: a between-subjects vignette experiment in which participants are told they are contributing health or financial data to a federated app. We randomize latency and failures, privacy disclosure (a differential-privacy $\epsilon$ statement versus a plain reconstruction-risk label versus none), and assurance (none versus a named accountable party). The outcome is a costly continuation choice rather than a Likert intention score. We will deliberately recruit beyond the young, highly educated samples that dominate this literature. The fitted $\beta$ values define a "calibrated" region inside the simulation grid.

The second component is a retention scheduler. Each round, the server estimates each client's departure risk from signals it already observes: that client's latency history, failures, and a sampled leakage proxy. It also estimates each client's irreplaceability, meaning how far the client's class distribution sits from what the remaining federation covers. Federated Learning with Non-IID Data shows that the Earth Mover's Distance between a client's label distribution and the population's bounds how far that client's local training pulls the model away. We reuse that distance as the irreplaceability measure.

The server then spends a fixed per-round budget on the clients with the highest product of departure risk and irreplaceability. It can shorten their local work, schedule them in reliable time slots, or project extra noise onto their leakiest update directions in the style of InvL-GNP. To keep costs manageable, it uses a low-rank randomized estimate rather than a full singular value decomposition, since the source paper concedes that full decompositions are too expensive on edge hardware. The scheduler's objective is expected full-population accuracy at the final round. This is the new ingredient: the levers act on who stays, not on how the missing updates are reweighted.

Baselines

The naive reference point is plain FedAvg (averaging client updates weighted by data size) run under risk-driven dropout with no correction. It sits alongside FedAvg with full participation, which serves as an upper bound. Together they show how much churn costs.

The strongest post-hoc competitor is Debiasing FedAvg from Debiasing Federated Learning with Correlated Client Participation. It is the method best designed for participation that is correlated rather than random. It reports 73.8% CIFAR-10 test accuracy under a five-round lockout, against 67.6% for FedAvg and 68.0% for FedVARP. We extend it naturally for churn: weights are frozen at the client's last observed frequency.

ReBaFL is the strongest competitor for protecting classes that go missing. It reports 68.59% on CIFAR-10 with two classes per client and 50% dropout, against 64.37% for FedAvg (Stabilizing and Improving Federated Learning with Non-IID Data and Client Dropout). FL-FDMS is the strongest substitution method. It reports no numeric CIFAR-10 accuracy on Lacuna, only that it approaches full participation under up to 70% dropout.

Two baselines lower perceived risk uniformly rather than selectively. The first adds uniform noise to every client's update, in the spirit of the untrusted-server private training of Private Federated Learning Without a Trusted Server. The second applies InvL-GNP to everyone. On STL10 with horizontal federated learning, InvL-GNP reports 84.8% accuracy against 83.1% for standard noise.

All methods get the same data partitions, model architecture, number of rounds, client count, and wall-clock and upload budget. Each method gets the same number of hyperparameter trials, tuned on a held-out setting of the dropout model. Beating the field means higher full-population accuracy and higher worst-class accuracy than the best of these methods, across the calibrated region of the $\beta$ grid, with no loss of privacy as measured by reconstruction error under a gradient-inversion attack.

Two ablations isolate the key ingredient. In the first, the scheduler spends its budget uniformly instead of targeting risk times irreplaceability. In the second, the scheduler is given the true departure probabilities instead of estimated ones, which separates the value of targeting from the cost of estimation error.

Why not just…?

Why not just use Debiasing FedAvg, which already handles correlated participation? This is the real threat. If risk-driven dropout turns out to be mostly temporary, with clients pausing and returning, then inverse-frequency reweighting could remove most of the bias and our scheduler would add little. The simulator therefore includes a parameter for how likely a departed client is to return, sweeping from fully temporary to fully permanent. We predict that Debiasing FedAvg matches our method when departures are temporary and falls behind as they become permanent. If it matches everywhere, the core claim is wrong and we will report that.

Why not just give everyone strong privacy protection, so perceived risk drops across the board? Uniform noise has a well-known accuracy cost, and the gap between InvL-GNP and uniform noise on STL10 already shows that where the noise goes matters. The deeper objection is that the calibration study may find that privacy disclosure barely moves continuation while failures and latency do. In that case uniform privacy spending buys no retention at all. The per-lever ablation reports which levers actually retain clients.

Why not just substitute a friend's update or use ReBaFL's prototypes for missing classes? Both methods assume that some similar client, or some stored class summary, stays available. Risk-driven churn tends to remove whole groups of sensitive-data clients together. Friend substitution then has no friend to use, and prototypes go stale because nothing refreshes them. We test this directly by varying how strongly risk-aversion correlates with label mix. If these methods hold up even under strong correlation, retention scheduling is unnecessary for the image tasks and its case rests on tabular settings.

Proposed Experiments

The simulated experiments use CIFAR-10 and Fashion-MNIST with two classes per client, matching ReBaFL's setting, and Dirichlet partitions, with 100 to 250 clients. We add HAR for realistic sensor clients. We also use the Obesity and Health Insurance tabular datasets from the untrusted-server private training work, where assigning higher risk-aversion to sensitive records is most defensible. Metrics are full-population and worst-class accuracy, retention curves over rounds, accuracy on the data distributions of departed clients, reconstruction error under a gradient-inversion attack, and scheduler compute per round.

The decisive check is a matched-marginal comparison. For every risk-driven run, we build an exogenous twin: a run in which the same clients drop out in the same rounds, but chosen at random rather than by risk. If existing methods lose the same accuracy in both conditions, endogeneity does not matter and the proposal's premise fails. If they lose markedly more under risk-driven dropout, and the retention scheduler closes that gap inside the calibrated $\beta$ region, the claim holds. A second check asks whether the scheduler's gains survive when its departure-risk estimates are noisy and the true $\beta$ values lie outside the region the scheduler was tuned for, since the behavioral link here is only weakly established.