Spending Privacy Budget Where the Satellites Are Blind

During a flood, the freely available satellite picture of a city goes stale between orbital passes. A radar satellite such as Sentinel-1 returns to the same spot only every 6 to 12 days, and cloud cover can blind optical satellites for longer. In those gaps, the most current information comes from sensitive ground sources: traffic cameras, drone video, and people's social-media photos of flooded streets. Differential privacy is a mathematical guarantee that a released result barely changes whether or not any one person's data was included. It lets agencies publish aggregates from these ground sources, such as the share of reports saying a road is impassable in each district. The catch is that the guarantee draws on a fixed, non-renewable "privacy budget," and every release spends some of it.

This note proposes that the budget should follow the satellite schedule. Spend almost nothing right after a clear satellite pass, when the public imagery already shows where the water is. Spend heavily as that imagery ages and before the next usable pass. Satellite ephemerides (the published predictions of where each satellite will be and when), cloud forecasts, and public imagery contain no personal data. A schedule computed only from them is therefore adaptive without costing any privacy. Existing dynamic allocators get their adaptivity by reading the private data or a tabular public proxy, and pay for it in leaked privacy or in domain mismatch.

Core Claim

In continual private release of ground-sensed disaster conditions, the right per-step budget is set by how uncertain the public-only estimate is at that step. That uncertainty is driven mostly by the time since the last usable satellite pass and the time until the next one. We claim that an allocation computed from this public uncertainty will beat the two families of schedules the privacy literature currently offers, at equal total privacy.

The first family, noise decay, says spend late. Yu et al. propose exponential and step decay under zero-concentrated differential privacy (zCDP), a variant with simple additive accounting. Hong et al. prove exponential decay is optimal for gradient descent under smoothness and Polyak-Łojasiewicz assumptions. The second family, frontloading, says spend early. Jiang et al. inject less noise when training loss is high, and Huang et al. add noise as updates stabilize.

The research file frames "spend late" versus "spend early" as an unresolved tension for disaster response. Our position is that both are the wrong shape for this problem. Both are tied to optimization time. In a monitoring stream, the value of private data is tied to what the public sensors cannot currently see, and that value is sawtooth-shaped across revisits. It is neither monotone nor tied to training progress.

Mechanism

Let $x_{k,t}$ be the true flood state of district $k$ in time window $t$, for example the fraction of roads that are impassable. A public-only estimator produces a variance $P_{k,t}$ for this state. That estimator is a simple state-space filter fed only by satellite passes, with process noise calibrated to how fast floods change. $P_{k,t}$ collapses at each clear pass and grows between passes.

The private ground aggregate is released with the Gaussian mechanism. Under zCDP, spending $\rho_{k,t}$ adds noise of variance $\Delta^2/(2\rho_{k,t})$, where $\Delta$ is the contribution-bounded sensitivity of one person's reports.

If each window is treated independently, the fused posterior variance is $(1/P_{k,t} + 2\rho_{k,t}/\Delta^2)^{-1}$. Minimizing urgency-weighted error under a total budget $\rho$ gives

$$\min_{\rho_{k,t}\ge 0}\ \sum_{k,t} \frac{w_{k,t}}{1/P_{k,t} + 2\rho_{k,t}/\Delta^2}\quad \text{s.t.}\quad \sum_{t}\rho_{k,t}\le \rho \ \text{for each } k,$$

where $w_{k,t}$ is a public urgency weight, such as population density or the response phase. Setting the derivative equal to a Lagrange multiplier $\lambda$ gives a water-filling solution:

$$\rho_{k,t} = \max\!\Big(0,\ \tfrac{\Delta^2}{2}\big(\sqrt{2w_{k,t}/(\lambda\Delta^2)} - 1/P_{k,t}\big)\Big).$$

Windows just after a clear pass, where $P_{k,t}$ is small, can receive zero budget. Stale windows receive the most. The full version propagates private information forward through the filter and is solved numerically. The schedule can also be re-planned online as cloud forecasts update.

Every quantity in the allocation ($P$, $w$, pass times, cloud forecasts) is public. The schedule is therefore post-processing of public data. Composition gives a total guarantee of $\rho$ per person per district, even for someone who reports in every window. Cross-district contributions are handled by per-person contribution bounding.

Two prior works play a supporting role. Wang and Zhou (PPSGD) supply the enabling idea that public signals can steer private allocation for free. They use a tiny public sample drawn from the same tabular distribution. We replace it with a physically distinct public sensor whose blind spots are exactly what the budget should cover. That sidesteps the public-to-private domain mismatch the file flags as a limitation of PPSGD. PrivateKube motivates treating budget as non-replenishable per data block, but it splits blocks by calendar time with no awareness of the physical sensing schedule.

Research Program

We will build three things.

The first is the allocator: the closed-form rule above and a numerical version that propagates information forward through the filter. Both will run under a zCDP accountant with per-person contribution bounding.

The second is a real evaluation stream built from the MediaEval 2018 Multimedia Satellite Task. It contains 11,070 timestamped Hurricane Harvey social-media posts labelled for road passability, together with WorldView-2 imagery. We treat the posts as the private ground stream, with the author as the privacy unit, and aggregate them by district and time window. Public satellite information comes from the Harvey Sentinel-1/2 and very-high-resolution stack used by Multi3Net, with acquisition timing and cloud flags taken from the actual scenes.

The third is a semi-synthetic camera stream, which lets us vary revisit gaps systematically. Virtual traffic cameras are placed on road networks. Their per-category detection error is calibrated from the submergence classes of Zhong et al., who report 89.29% mAP without privacy. Flood-extent change magnitudes are calibrated from the multi-temporal events in Kuro Siwo. Revisit gaps are swept from daily to 12 days, and cloud outage rates from zero to multi-week.

We will also analyze when the gain over uniform allocation should vanish: when revisits are frequent relative to how fast the flood changes, $P_{k,t}$ is nearly constant and water-filling reduces to a uniform split. This gives a falsifiable prediction about where the method matters.

Baselines

The naive reference points bracket the problem. At one end is public-only estimation, which spends zero budget and uses satellites alone. The closest reported figure is Multi3Net's 59.7% mIoU for flooded buildings from freely available Sentinel-1 and Sentinel-2 alone, against 75.3% with very-high-resolution imagery (Multi3Net). That is a different task, but it shows how much public-only sensing leaves on the table. At the other end is the non-private fusion of ground and satellite data, which serves as the upper bound. Every private method must land between these two. Beating public-only by a clear margin at realistic budgets is the minimum bar.

The uniform split divides $\rho$ equally across windows and feeds the same fusion filter. Holding the estimator fixed isolates the allocation.

The strongest existing method is optimization-time dynamic scheduling. We will run exponential decay tuned as in Yu et al. and Hong et al., and a frontloaded mirror image in the spirit of Jiang et al.. Yu et al. report accuracy gains over uniform noise of 1.0–1.6% on MNIST and 4% on CIFAR-10 transfer. PPSGD reports mean squared error of 0.1027 versus 0.1253 for the best private baseline on cadata at $\epsilon=0.3$. No reported result was found on Lacuna for any private release on MediaEval, Harvey, or traffic-camera flood data. The comparison here is therefore new rather than a replication.

A PrivateKube-style baseline releases a fixed budget per calendar day, following Luo et al..

A private-adaptive baseline spends a sub-budget detecting change in the private stream and releases only when change is flagged. This is the closest analogue to data-driven allocators like ALPINE. It is run with the same total $\rho$, including what the detector spends.

All schedules get identical tuning budgets. Hyperparameters are chosen on a held-out event or held-out simulation seed, never on the test stream.

Two ablations isolate the key ingredient. A simple linear ramp in time-since-last-pass tests whether the variance-based rule matters beyond the revisit signal itself. A variant using nominal orbits without cloud information tests how much cloud forecasts contribute.

"Beating" means lower urgency-weighted error, and higher road-passability accuracy against the non-private fused estimate, at every budget in a sweep from $\rho=0.05$ to $\rho=2$. The gain must also be statistically separable across simulation seeds and districts.

Why not just…?

Why not just split the budget uniformly and let the fusion filter do the rest? This is the default and it will be competitive when revisits are frequent. Our own analysis predicts equivalence in that regime. The semi-synthetic sweep over revisit gaps is designed to show where uniform allocation stops being good enough, rather than to assume it fails. If uniform matches us at the real 6–12 day Sentinel-1 gaps, the claim is wrong and we will report that.

Why not just use the published decay or frontloading schedules? Their justification is optimization-specific. Hong et al.'s optimality proof concerns how early gradient noise fades during descent under Polyak-Łojasiewicz assumptions. In a monitoring stream, nothing "converges," and the next satellite pass resets what private data can add. We expect decay to waste budget right after passes that happen late in the horizon, and frontloading to starve the stale windows before later passes. The head-to-head comparison at matched $\rho$ settles this directly.

Why not just make allocation adaptive to the private data itself, as ALPINE or change detection does? Data-driven adaptivity can react to surprises the public sensors miss, such as a levee breach between passes, so this is a real threat. Its cost is that the detector consumes budget, and its decisions must be accounted for. We run it at equal total budget. We also test a hybrid in which a small private change-detection sub-budget raises $w_{k,t}$. If the pure private-adaptive method wins, the public-staleness signal is not sufficient, and the hybrid becomes the recommendation.

Why not just use the time-since-last-pass ramp and skip the variance model? If the ramp matches water-filling, the useful finding is simply that revisit timing is the variable that matters, and the simpler rule should be deployed. We include it as an ablation precisely so that the proposal's value does not rest on the formalism.

Why not just blur faces and skip differential privacy? Blurring gives no formal guarantee against membership inference. The aggregate-release setting here also does not need raw pixels to leave the device. This alternative is out of scope rather than a competitor on the utility question.

Decisive Checks

The central experiment plots error against budget on the MediaEval Harvey stream, using real pass and cloud timing, for all schedules sharing one fusion filter. The claim survives if water-filling sits below uniform, decay, frontloading, and the calendar-block baseline across the budget sweep. It must also recover a substantial fraction of the gap between public-only and non-private fusion at moderate budgets.

The second check is the revisit-gap sweep on the semi-synthetic camera stream. The gain over uniform should grow with the gap and with cloud outage rate, and shrink toward zero at daily revisit. If the gain is flat across gaps, the mechanism is not what we think it is.

The third check is the private-adaptive comparison under an injected surprise event between passes. This measures the price of relying only on public signals and determines whether the hybrid is needed.