Consent-Governed Federated Personalization with a No-Harm Check

Federated learning trains one model across many devices, such as phones, wearables and home sensors, without moving anyone's raw data off their device. It has become the default way to build health models for older adults and other people whose data is legally and ethically sensitive. In current systems, though, the person behind each device is a passive data source. A server decides who trains in each round. An algorithm decides how much personalization each person receives. Nobody checks whether the personalized model actually serves that person better than a generic one would. Work on the "active participation model" in eldercare argues the opposite stance. People should choose what they contribute and on what terms. Systems built around passive monitoring breed dependency and withdrawal (Promises and Challenges of Ambient Assisted Living Systems; The missing ones).

This note proposes to make that stance an operational part of federated training. Each person decides which of their sensor streams, if any, may be used to train the shared model. Before choosing, their device shows them an honest, locally computed estimate of whether each option would actually help them. The practical question is whether a federated system can let people choose, keep them from being made worse off by personalization, and still train well when the people who opt in are not a random sample.

Core Claim

Participatory Personalization in Classification supplies the key concept, "worsenalization." It means that personalizing a model on someone's attributes can leave that person with higher error than a generic model. That paper lets individuals opt into disclosing attributes at prediction time. It works on a centralized dataset and guarantees its participatory systems never do worse than the generic model.

The federated healthcare line never checks for this failure. FedHealth personalizes by freezing a shared feature extractor and aligning each user's top layers to their own data. FedHealth 2 weights each client's aggregation by the similarity of their batch-normalization statistics. Both report average accuracy only and assume personalization always helps.

The proposal makes three claims:

  1. Worsenalization exists in federated personalization. A measurable fraction of individual clients in these systems end up worse off than under a generic federated model.
  2. Opting in is not random. When participation is voluntary, the people who opt out tend to differ systematically from those who opt in. This biases the shared model against non-participants, a problem that uniform-sampling analyses ignore.
  3. The fix is a per-person no-harm check with choice. Train one model that works under any subset of disclosed sensors. Let each person choose among options that pass a local statistical test for improvement over the generic model. This recovers most of FedHealth 2's accuracy while cutting the worsenalized fraction to near zero.

The honest scope is this. If a simple "pick whichever of global or local validates best" rule does just as well, the contribution shrinks to the measurement study in the first two claims. We state that outcome in advance.

Research Program

Shared model that tolerates missing sensors. The first component is one federated model that stays usable under any subset of sensor channels. During local training we randomly mask input channels, so the model learns to work with whatever a person chooses to disclose. We keep local batch normalization, following FedHealth 2, to absorb differences in sensor calibration. Clients that disclose the same channels are aggregated with similarity weights. This reuses FedHealth 2's method; the change is that aggregation is conditioned on each person's disclosure choice rather than on all data.

Local no-harm check. The second component runs entirely on the device. For each disclosure option $d$, including "contribute nothing and receive the generic model," the device compares option $d$ against the generic model $g$ on the person's own held-out labeled windows $V_i$. It computes a one-sided lower confidence bound on the accuracy gain:

$$\mathrm{LCB}_i(d) = \hat{\Delta}_i(d) - z_{1-\alpha}\,\hat{\sigma}_i(d)/\sqrt{|V_i|}.$$

Here $\hat{\Delta}_i(d)$ is the mean paired accuracy difference on $V_i$, $\hat{\sigma}_i(d)$ is its standard deviation, and $\alpha$ is the tolerated risk of harm. The $\sqrt{|V_i|}$ term is the usual standard-error scaling. Only options with $\mathrm{LCB}_i(d) > 0$ are presented as "likely to help you." Everything else defaults to the generic model. This carries the gain-verified disclosure idea of Participatory Personalization into federated training. Unlike Active Federated Learning, which has clients upload their loss values so the server can pick who trains, nothing about the check leaves the device. The person, not the server, decides.

Simulated self-selection. The third component simulates who opts in. We let the probability of opting in depend on each client's data, for example on the share of low-intensity activities as a stand-in for frailty. We then measure the harm to both opt-ins and opt-outs as the overall opt-in rate varies.

This borrows a question from Participant: A New Concept. That study's community simulation shows service failures falling non-linearly as the participant rate rises: 987 aborted requests at 10% participation, 99 at 25%, and 5 at 60%. We ask whether federated personalization has an analogous threshold below which non-participants are poorly served.

Scope. We do not claim this delivers empowerment in the social-science sense. The illusion of artificial inclusion warns against simulating human agency, and our simulated opt-in behavior is a stress test, not a substitute for people. A small co-design pilot on how to present the no-harm check to older users is a follow-on step that is out of scope here.

Baselines

Uncollaborative references. These are purely local training and a single generic FedAvg model. On PAMAP with 20 Dirichlet-split clients, FedHealth 2 reports 69.07% accuracy for local training, 67.58% for FedAvg, and 67.56% for FedBN. The generic FedAvg model also serves as the reference point that defines worsenalization for each client.

Strongest existing method. This is FedHealth 2, at 81.14% average accuracy on the same PAMAP split. We rerun it under our code with the same partition seeds, model, rounds and hyperparameter search budget. On UCI HAR, FedHealth reports 99.4% average over five held-out users. That benchmark is near ceiling, so we use it only as a sanity check.

We count the proposal as beating FedHealth 2 if both of the following hold:

  1. The fraction of worsenalized clients is significantly lower across seeds.
  2. Mean accuracy is within two points of FedHealth 2, with no drop in the worst client's gain over generic.

We do not expect to beat its mean accuracy. Trading a little average accuracy for per-person safety is the claim.

Ablations. Two ablations isolate the key ingredients:

  • Validation-picker. Each client simply picks whichever of the generic, local or FedHealth 2 model scores best on its validation set, with no confidence bound. This tests whether the statistical no-harm check matters.
  • No channel masking. All disclosure options are tried without training the model to tolerate missing sensors. This tests whether one model can serve all disclosure choices.

Sampling baseline. Every method is additionally run with Active Federated Learning client sampling, which the paper reports raised Reddit AUC from 0.552 to 0.578. This shows whether our findings hold when the server, rather than the person, selects participants.

Why not just…?

Why not just let each client keep whichever model validates best? This is the real threat. With many labeled windows per client it may be sufficient. With few labels, which is the realistic case for older adults, picking the maximum of noisy validation scores tends to overstate the winner. That produces exactly the hidden worsenalization we are measuring. The validation-picker ablation settles this directly. We shrink $|V_i|$ from hundreds of windows down to about twenty and report where, if anywhere, the confidence bound starts to matter. If it never matters, we report that the simpler rule is enough.

Why not just add an opt-out flag to server-side selection such as Active Federated Learning? An opt-out flag gives people a veto but not the information needed to use it well. It also ignores that opting out is correlated with who a person is. Active Federated Learning's design also requires uploading raw loss values, which that paper itself flags as a privacy leak. The self-selection experiment tests whether correlated opt-out harms non-participants even with active sampling. If harm appears only without active sampling, the simpler fix wins on that question.

Why not just pay people to participate using Shapley-based reciprocity? You Get What You Give shows its payment mechanism, M^Shap, raises data sharing 9.51x on MNIST. However, it rests on smoothness assumptions about how accuracy and effort costs scale, and it requires costly contribution estimation. More basically, payment does not tell a person whether personalization will hurt them. The two approaches are complementary rather than competing. We therefore treat reciprocity as out of scope rather than claim superiority over it.

Proposed Experiments

Measurement study. On PAMAP, we rerun FedAvg, FedBN and FedHealth 2 over at least five partition seeds and report per-client accuracy relative to the generic model. The worsenalized fraction is the headline number. No per-client figure is reported in the file for any of these methods, so this measurement is itself new. Disclosure options are subsets of PAMAP's body-worn sensor groups. We run UCI HAR as a lower-difficulty check.

Decisive checks. Three results decide the claims:

  1. Worsenalization. Does FedHealth 2 worsenalize a non-trivial share of clients, and does the no-harm check eliminate it at a mean-accuracy cost of at most two points?
  2. Label scarcity. Does the no-harm check beat the validation-picker as per-client labels shrink?
  3. Opt-in threshold. When opt-in rates are swept from 10% to 100% with data-correlated opt-in, does accuracy for non-participants degrade non-linearly below some threshold, as the eldercare simulation found for service failures?

A clean negative on the first check, meaning federated personalization rarely harms anyone, would be informative in its own right and would redirect the work toward the self-selection question alone.