Randomized Defaults as Built-In Instruments for Unbiased Federated Learning

Federated learning trains one shared model across many devices or hospitals. Each participant trains locally and sends back only a model update, never its raw data. These systems quietly depend on who agrees to take part. If people who opt out differ systematically from those who stay in, for example because the model already serves them badly, the shared model drifts toward the participants and grows confidently wrong for everyone else. Current fixes try to undo this bias after the fact by reweighting the updates that do arrive. That only works if the server can estimate each client's chance of participating, which in turn requires side information whose validity cannot be checked from the data.

Behavioral economics offers a lever the federated learning literature has ignored. The default, meaning what happens if a user does nothing, strongly shifts participation without changing anyone's underlying data. Our proposal is not to use defaults to push participation up. It is to assign defaults at random: half the clients see "participate unless you decline," half see "decline unless you join." Because the default changes whether a client participates but is independent of that client's data, the random assignment acts as a designed instrument, a source of variation that lets the server identify and correct the selection mechanism without guessing.

Core Claim

Randomizing participation defaults (and, analogously, privacy-tier defaults) across clients makes non-random dropout correctable from the updates the server already receives. Correction then becomes possible even when the side signals that existing methods rely on are weak or missing. Fixed defaults, however strong, cannot provide this. They raise the participation rate, but the clients who still opt out are exactly the ones with the strongest reasons, and their missingness stays unidentified.

The mechanism is a nonresponse-instrument estimating equation. Let $Z_k$ be the default randomly assigned to client $k$, $R_k\in\{0,1\}$ whether it contributes an update, and $s_k$ a client-level statistic that is visible only when $R_k=1$, such as the current global model's local loss, which arrives with the update. We model the participation probability as $\pi_\phi(Z_k, s_k)$ with parameters $\phi$ and estimate $\phi$ from

$$\mathbb{E}\Big[\Big(\tfrac{R_k}{\pi_\phi(Z_k,s_k)}-1\Big)h(Z_k)\Big]=0,$$

where $h$ is any vector of functions of the assigned default. The left side is computable because the ratio is nonzero only for participants. It identifies $\phi$ because $Z$ is independent of $s$ by construction while still moving $R$. The server then aggregates with inverse-propensity weights:

$$w_{t+1}=w_t+\sum_{k:\,R_k=1}\frac{n_k}{N\,\hat\pi_k}\,\Delta w^k_{t+1}.$$

Here $w_t$ is the global model at round $t$, $n_k$ is client $k$'s data size, $N$ is the total data size across all clients, $\hat\pi_k$ is the estimated participation probability, and $\Delta w^k_{t+1}$ is client $k$'s update. Each additional default arm, such as a different default privacy tier or prompt timing, adds instrument values and so supports a richer participation model.

Relation to Prior Work

Two papers motivate the gap by correcting selection only after it happens. FLOSS models client opt-outs as missing-not-at-random, meaning missingness that depends on the unobserved outcome. It corrects them with inverse-probability weighting identified through hardware "shadow variables" and voluntary satisfaction prompts, and it treats the opt-out prompt as a fixed given. HDMI fits a distributed two-stage Heckman selection model on electronic health record features. It needs a domain-engineered instrument and becomes unstable at extreme missingness rates. We reuse their weighting machinery. The difference is that we manufacture the instrument through interface design, so it is valid by randomization rather than by assumption.

FeO2/FedHDP shows that a small set of clients opting out of differential privacy (calibrated noise added to protect individuals) can anchor the global model. It treats who opts out as exogenous. If the self-selected non-private clients are unrepresentative, for instance beta testers, the anchor is biased. Randomized default tiers let us test and correct this.

The behavioral evidence constrains how large the instrument can be. In Cherep et al., human default adoption in a sequential choice task is 52.6% without a default and 89.9% with one, which is a strong first stage. LLM agents reach 97.6–99.8%, which matters if edge clients become agents. Manisterski et al. show that even capable users almost never change supplied defaults. Opt-in versus opt-out framing studies such as GAMEPAD confirm large participation shifts in consent settings. Chen et al. warn that opaque defaults become covert steering. Our design keeps every default disclosed and reversible, and in both arms the user retains full choice.

Research Program

We will build a federated simulator on Flower, as used by FLOSS, or on decentralizepy with LEAF datasets. Each client gets a behavioral participation model, $P(R=1\mid Z,s)=\sigma(\alpha+\beta Z+\gamma s)$. In this model:

  • $\beta$ is the default effect. We calibrate it so that adoption moves from roughly 53% to 90% (the human figures in Cherep et al.). We also sweep it, including the near-deterministic agent regime.
  • $\gamma$ couples dropout to how badly the model serves the client.
  • A second mechanism ties dropout to group membership, such as low-data users.

We transfer these numbers from a different task, so we treat them as a calibration point and not as ground truth.

On top of the simulator we will implement three things. The first is the estimating equation above, with the server solving for $\phi$ each round. The second is a design analysis of the identification–participation trade-off. Assigning the weaker default to a fraction $q$ of clients costs updates but sharpens identification, and we will derive and empirically trace the variance-minimizing $q$. The third is an extension to privacy tiers. Here $Z$ is the default differential-privacy setting, $R$ is opting out of differential privacy, and the estimated propensities reweight FedHDP's non-private anchor.

Baselines

The naive reference is uncorrected FedAvg, the size-weighted average of client updates, under non-random dropout. It is the do-nothing system, and FLOSS reports it losing 8–10% accuracy relative to complete participation on its simulated binary task.

The strongest existing method is FLOSS's inverse-probability weighting with shadow variables and satisfaction prompts. On that same task, FLOSS reports recovering the loss to closely match an oracle that knows the true participation probabilities. We run it on identical client populations, rounds, and local-training budgets, and give it the same hyperparameter search. We vary the informativeness of its shadow variable from strong to null, and the response rate of its satisfaction prompt.

A second correction baseline is a federated Heckman model in the style of HDMI, given the best available engineered instrument. HDMI reports under 5% bias versus 5–15% for local imputation in its electronic health record simulations.

The upper references are the oracle-propensity model and a no-missing-data model.

The key ablation is a fixed opt-out default for all clients, tuned to the same average participation as our randomized design. This isolates whether randomization, rather than simply having more participants, does the work.

For FEMNIST and Sentiment140 under non-random client dropout, no reported result was found on Lacuna for federated accuracy or bias, so all numbers on those datasets will be produced in-house under shared settings. Beating the baselines means two things. First, we must match the oracle within seed noise in the regimes where FLOSS succeeds. Second, we must stay near the oracle where FLOSS's shadow variable is uninformative and it falls back toward uncorrected FedAvg.

Why not just…?

Why not just set the participate-by-default option for everyone and maximize participation? This is the most serious threat. If 90% of clients comply, the remaining missingness may be too small to matter. However, the residual decliners are the clients with the strongest reasons to leave, so their missingness is the most selective, and with a single default it is not identifiable at all. The fixed-default ablation settles this directly. We will report the region of default strength $\beta$ and dropout coupling $\gamma$ in which a fixed default is good enough. If that region covers realistic calibrations, we will say the randomized design is unnecessary there.

Why not just use FLOSS's shadow variables? When device metadata truly predicts dropout and users answer satisfaction prompts, FLOSS should be about as good as our method, and we expect to tie it there. Its weakness is that shadow-variable validity is untestable, and prompts are themselves subject to non-response. The sweep over shadow-variable strength shows how much of the oracle's performance each method retains as that information disappears.

Why not just pay clients through contract-based incentives? Randomized payments would also serve as an instrument, and that is a fair point. The difference is cost: defaults are free. The behavioral evidence (Manisterski et al., Thaler's Behavioral Economics) also suggests friction outweighs small payments. We have no data on payment elasticity in federated settings, so we include randomized incentives as a simulated arm under assumed elasticities and do not claim to beat them in practice.

Why not just enforce uniform differential privacy and remove the privacy choice? FeO2 reports up to a 9.27% gain over uniform differential-privacy FedAvg when 5% of clients opt out. Removing the choice therefore gives up real accuracy. Our privacy-tier experiment asks whether that gain survives once opt-outs are self-selected rather than random.

Proposed Experiments

We run the simulator on LEAF's FEMNIST and Sentiment140, LEAF Synthetic, and a FLOSS-style binary task. Metrics are global accuracy on all clients' test data, including non-participants. We also report 10th-percentile client accuracy, since LEAF notes that lower percentiles collapse while the median looks fine, along with bias of the estimated population loss, seed-to-seed variance, and participation rate.

The first decisive check is null-shadow robustness. With an uninformative shadow variable, we expect our estimator to stay near the oracle while FLOSS degrades.

The second is the fixed-versus-randomized comparison at equal participation. If the fixed default matches us across realistic $\beta$ and $\gamma$, the claim fails.

The third traces the cost curve over $q$ to locate the optimal split.

The fourth stress-tests the core assumption by letting the default leak into the data, for example by making opt-in-arm clients train on systematically different data. This violates the requirement that the default be independent of the data, and we measure how quickly the correction breaks.

The fifth is the privacy-tier variant, testing whether FedHDP's anchor is biased under self-selected non-private clients and whether default-based propensities repair it.

The sixth is an agent-client regime with near-total compliance (97–99%). It asks whether the very strong instrument that agents provide comes with selection that carries no information about the data.