Indoor positioning systems guess where a phone or robot is from the radio signals it hears. They work by learning a "fingerprint" map from signal patterns to coordinates, and that map goes stale whenever the building changes: a forklift moves, a shelf is rebuilt, an access point is relocated. Federated learning, in which many sites train one shared model by sending model updates instead of raw data, is a natural way to keep these maps current across a campus or a set of warehouses. Those updates still leak information, though. Attackers can partly reconstruct training inputs from them, and for a localization model a reconstructed input is effectively a user's position. The standard remedy is a uniform privacy defense: clip every update to a fixed size and add noise.
This note argues that the remedy fails unevenly, and that the unevenness follows drift. A zone whose radio environment just changed sends large updates that look unlike everyone else's. Those updates are the easiest to invert and the most damaged by clipping and noise. Uniform privacy therefore abandons the zone that most needs to adapt. We propose to measure this effect on real radio benchmarks, using a leakage metric expressed in meters. We then propose a constrained federated training rule that caps how much worse each zone may become because of privacy, relative to the same system without privacy. The rule pays for that cap with extra local computation and does not loosen the privacy guarantee.
Core Claim
Under staggered temporal drift, the clients that are currently adapting have three properties at once. They are the most exposed to gradient inversion, the most degraded by uniform clipping and noise, and the most hidden by server-side averaging. The resulting disparity in localization error across zones should be treated as a fairness constraint inside training, not fixed after training. Each zone's private-model error is bounded relative to its own non-private reference. Per-zone Lagrange multipliers, which are adjustable penalty weights that rise when a constraint is violated, set how hard that zone works locally before its clipped update is sent. Clip norm and noise scale stay fixed.
The closest existing work combines fairness, differential privacy and federated learning on static data. PUFFLE adds a demographic-parity penalty to DP-SGD on CelebA, Dutch and ACS Income. PrivFairFL computes fairness statistics under secure multiparty computation. FedFair solves equal-opportunity constraints by alternating gradient projection. All three define groups by a fixed demographic attribute on a stationary distribution. Here the groups are physical zones, and which zone is disadvantaged changes over time with the environment. The fairness target is also different: none of them asks whether the privacy mechanism itself creates the disparity.
On the drift side, FedDrift handles staggered drift by clustering clients. Its own authors flag that small clusters raise leakage risk, which is one of the hazards we measure. FedAMP for indoor localization personalizes models across static spatial partitions of UJIIndoorLoc, with no time dimension and no leakage analysis.
Research Program
The work has three parts.
A radio-native leakage audit. Vision inversion studies score attacks with image metrics such as LPIPS, PSNR and SSIM. Huang et al. note that these metrics may not transfer to signal vectors, and for localization they are beside the point. We will run optimization-based gradient matching against the localization model, with and without a learned signal prior in the spirit of generative-prior inversion. We then pass each reconstructed signal vector through the current global model and report the distance between the decoded position and the true position. This "position leak" is measured in meters.
Over multiple rounds we also run the user-level linking attack of Wang et al. to test whether trajectories can be recovered. A third test asks whether an attacker can tell from update statistics alone which zone has drifted. That would reveal the physical layout change itself.
The hypothesis is that drift-phase updates leak more than steady-state updates. The mechanism comes from Wu et al., who show that inversion succeeds when per-sample updates are weakly entangled. Novel post-drift signals should produce exactly such weakly entangled updates.
A disparity audit of standard defenses. We will apply three defenses and record per-zone error over time, separately for zones that drifted and zones that did not. The defenses are DP-SGD with a fixed clip, gradient pruning, and the pruning-plus-encoding combination from Huang et al.. The quantity of interest is each zone's excess error over its non-private reference, tracked as drift happens.
Constrained private adaptation. The method comes from Balancing Act. That paper constrains each subgroup's accuracy drop in a pruned network relative to the dense network. We transfer the idea, with privacy playing the role of pruning:
$$\min_\theta \; \frac{1}{Z}\sum_{z=1}^{Z} E_z(\theta) \quad \text{s.t.} \quad E_z(\theta) - E_z^{\text{ref}} \le \epsilon \;\; \forall z,$$
Here $E_z(\theta)$ is the mean positioning error of the private model in zone $z$, and $Z$ is the number of zones. $E_z^{\text{ref}}$ is the error of a non-private model trained on the same rounds. In practice we estimate it with a lagged proxy: the zone's own error just before drift, plus a drift allowance. $\epsilon$ is the tolerated excess error in meters.
Each zone's multiplier $\mu_z$ grows when its constraint is violated. The multiplier sets the client's number of local steps and its weight on recent samples. It never changes the clip norm or noise scale, so the formal guarantee is identical to that of the fixed-clip DP-SGD baseline.
Zone errors are computed on each client's small held-out labeled set. They are reported through noised secure aggregation, and that cost is charged to the privacy budget.
Constrained optimization of this kind is known to oscillate (benchmarking study). We will therefore damp multiplier updates and report how stable the multipliers are.
Baselines
The naive reference is FedAvg with fixed-clip DP-SGD at a matched privacy budget. This is the "do nothing about fairness" system, and the audit predicts it produces the drift-tracking disparity. The same configuration without privacy supplies $E_z^{\text{ref}}$ and the upper bound on utility.
The strongest privacy competitor is the defense combination of Huang et al.: gradient pruning at 0.9, intra-sample InstaHide encoding, and batch size 32. It reports LPIPS 0.73 against 0.45 undefended, at 86.10% CIFAR-10 accuracy against 93.37%. No reported result was found on Lacuna for this defense on radio fingerprints, so we will rerun it on our data.
The strongest fairness competitor is PUFFLE. It reports up to 75% unfairness reduction with at most 17% utility loss on tabular and image data. We will adapt its penalty so that groups are zones and the disparity is in error, and give it the same tuning budget for its penalty schedule.
The strongest drift competitor is FedDrift, which reports 64.8% against 58.8% for the best baseline on FMoW. We will run it under the same DP budget, accounting for the smaller effective aggregation size inside its clusters.
For centralized adaptation quality, 5G-DIL reaches 0.261 m mean absolute error on the Fraunhofer 5G dataset with 50 to 200 exemplars. We treat it as a non-private, non-federated ceiling.
The key ablation keeps everything in our method but sets the multipliers to a constant. This isolates whether constraint-driven allocation of local work, rather than simply doing more local work, closes the gap. A second ablation lets the multipliers raise the clip norm instead of the local work, at an honestly re-accounted privacy cost.
All methods share data splits, model architecture, number of rounds, client compute ceiling and privacy budget. Each gets an equal hyperparameter search.
Our method beats the baselines if, at matched privacy budget and matched position leak, it reduces the worst-zone excess error during drift. The target is to bring it within $\epsilon$ while keeping mean error no worse than PUFFLE-style penalization.
Why not just…?
Why not just use differential signal features so that drift disappears? Differencing the signal strengths of nearby access points cancels a uniform attenuation, such as a crowd that dims every path by the same amount (Wang et al.). It cannot undo structural changes that reroute individual reflection paths (Zhao et al.). It also does not apply to channel impulse responses, where 5G-DIL reports error climbing once shifts exceed 0.6 m. We will include differential features as an input variant for every method. If drift-induced disparity vanishes under them on UJIIndoorLoc, we will say so and restrict the claim to channel-impulse-response settings.
Why not just give drifting clients a larger clip norm or adaptive clipping? This is a real threat. It is the cheapest fix and might close the error gap. It does so by spending more privacy budget on the very clients we expect to leak most. It also lets an observer read drift status from the size of each client's updates. Our second ablation runs exactly this with honest accounting, and the leakage audit measures both its position leak and its layout leak. If it matches our method at equal accounted budget and equal leak, our mechanism adds nothing.
Why not just let the drifting zone fine-tune locally and skip federation during adaptation? This is also a real threat. 5G-DIL shows that a few hundred exemplars can match full retraining, and keeping data local leaks nothing. Its weaknesses are that it assumes the zone holds enough fresh labeled data and that it stores raw exemplars (ConSense flags that as a privacy risk of its own). We will sweep the number of fresh labels per drifting zone. If local-only fine-tuning wins at realistic label counts, the honest conclusion is that drifting clients should temporarily leave the federation, and we will report that as the finding.
Why not just fix fairness after training by recalibrating per zone? Woodworth et al. prove that post-hoc adjustment is suboptimal because it cannot recover features that training discarded. In federated settings it can also be very expensive: PrivFairFL's post-processing takes about 28 hours for 100,000 samples (PrivFairFL). We still include per-zone bias correction of the private model's position outputs as a cheap baseline.
Proposed Experiments
UJIIndoorLoc. Its validation set was collected months later with different users and devices, which provides genuine temporal drift. Clients are the building-floor partitions. The reported centralized reference is a recurrent network's 8.54 m 3D error (UJIIndoorLoc study). To create staggered drift, we expose some clients to post-gap data several rounds earlier than others.
Fraunhofer 5G industrial dataset. The deterministic sites DS1 and DS2 act as stable zones. The realistic sites RS3 to RS5, which have moving vehicles and platforms, act as drifting zones, and they arrive in staggered order. No federated result is reported on Lacuna for this dataset.
WiAR. This Wi-Fi activity-recognition dataset is a secondary check that the leakage and disparity effects are not specific to localization. ConSense's 89.85% centralized accuracy serves as the reference.
Metrics. We report mean and worst-zone positioning error over time, excess error relative to the non-private reference, position leak in meters, trajectory-linking success, drift-zone identifiability from update statistics, privacy budget, and client computation.
Decisive checks. The first check is whether drifting clients' updates are measurably more invertible than stable clients' updates at the same round. If they are not, the premise linking leakage to drift fails, and the proposal reduces to a fairness-under-DP study. The second check is whether fixed-clip DP-SGD's disparity concentrates in recently drifted zones and decays as they settle. The third check is whether our constrained rule beats both the adaptive-clipping ablation and local-only fine-tuning at equal accounted privacy and equal position leak.
References
- PUFFLE: Balancing Privacy, Utility, and Fairness in Federated Learning
- PrivFairFL: Privacy-Preserving Group Fairness in Federated Learning
- Training Fair Models in Federated Learning without Data Privacy Infringement
- Federated Learning under Distributed Concept Drift
- Personalized Federated Learning over non-IID Data for Indoor Localization
- Evaluating Gradient Inversion Attacks and Defenses in Federated Learning
- Gradient Inversion with Generative Image Prior
- Beyond Inferring Class Representatives: User-Level Privacy Leakage From Federated Learning
- Concealing Sensitive Samples against Gradient Leakage in Federated Learning
- Balancing Act: Constraining Disparate Impact in Sparse Models
- Benchmarking Stochastic Approximation Algorithms for Fairness-Constrained Training of Deep Neural Networks
- 5G-DIL: Domain Incremental Learning with Similarity-Aware Sampling for Dynamic 5G Indoor Localization
- Differential radio map-based robust indoor localization
- Subspace-Based Super-Resolution Sensing for Bi-Static ISAC With Clock Asynchronism
- ConSense: Continually Sensing Human Activity with WiFi via Growing and Picking
- Learning Non-Discriminatory Predictors
- On the Use and Construction of Wi-Fi Fingerprint Databases for Large-Scale Multi-Building and Multi-Floor Indoor Localization: A Case Study of the UJIIndoorLoc Database