Wide-field astronomy has a well-tested answer to a problem that differential privacy handles poorly. A sky monitor cannot watch every patch of sky at full sensitivity all the time, so it runs a cheap, coarse, always-on "scout" and spends expensive, precise "sniper" observations only where and when the scout raises an alarm. We propose bringing this two-tier structure into private stream monitoring. Differential privacy (DP) is a guarantee that any one person's data barely changes the output, enforced by adding calibrated noise under a limited total "privacy budget" $\epsilon$. Picture a city-wide stream of location check-ins in which we want to detect and localize a sudden local surge, such as a crowd forming or an outbreak hotspot. When each person's data spans many time steps, the budget must be divided over time. Current methods either split it evenly or spend it on a single flat stream. Our design runs a low-budget private change detector on coarse spatial cells. When its noisy alarm fires, it releases a short, high-budget, fine-grained count over only the flagged region. A hard cap on the number of alarms per window makes the worst-case privacy cost fixed and simple to account for.
The core claim is that, for sparse, bursty, spatially concentrated events, putting the budget at the time and place of an alarm should localize events faster and more accurately than any static split at the same worst-case privacy cost. The intuition is that a burst lasting $L$ steps inside a privacy window of $w$ steps gets about $w/L$ times less noise per step from the sniper than from a uniform schedule. Equally important, we say up front when the idea should fail. If events are large and diffuse, or if privacy is defined per event rather than per person over time, a uniform per-cell detector should be just as good, and the experiments are built to show this.
Where the pieces come from
The astronomy direction supplies the architecture and the statistical regime, not the privacy need. Astronomy data is public, and the file contains no prior work joining the two fields. Three facts motivate the design. First, the proposed SWGO ground array is explicitly planned as an always-on scout whose alerts trigger deep follow-up by narrow-field Cherenkov telescopes, and it is simulated to outperform untriggered deep observation for roughly hour-long active-galaxy flares (The Sub-TeV transient Gamma-Ray sky). Second, the LOFT Wide Field Monitor broadcasts about 1-arcmin positions within 30 s for roughly 100 gamma-ray bursts per year from a field of more than 3.7 sr (The LOFT wide field monitor). Third, the lobster-eye pathfinder LEIA shows the realistic difficulties: heavy background, 10–15 usable minutes per orbit, and day-scale latency (LEIA onboard SATech-01). These systems describe the setting we target: Poisson counts against a large background, signals concentrated in a few spatial cells, gaps in coverage, and clustered arrivals. Astronomy also gives us a public simulator with known ground truth. We use it to stress-test the mechanism, not to claim that sky data needs privacy.
The privacy direction supplies the components and the constraints. Sequential Change Detection With Differential Privacy adds Laplace noise to the running CUSUM statistic and to its threshold. CUSUM is a cumulative log-likelihood sum that alarms when it crosses a threshold. The noisy version matches non-private detection delay to first order, at constant cost per step. The authors name multi-sensor networks, where only some streams change, as future work, and that gap is exactly ours. We reuse DP-CUSUM as the scout. Single and Multiple Change-Point Detection with Differential Privacy gives a sliding-window alternative that DP-CUSUM already beats on delay, so we treat it as a secondary baseline. Differentially Private Continual Release with Relative Error is the closest relative. Its "monitor-and-update" scheme watches a statistic with a noisy threshold test and spends budget only when the test fires. That work is limited to non-negative inputs and one released quantity. It does not change spatial resolution when triggered, and it proves polynomial lower bounds when the stream adapts to past releases. We take the trigger-then-spend logic and add the change in resolution, and we must keep alarm-driven follow-up from becoming the kind of adaptive feedback those lower bounds cover.
Two further papers constrain the design. LAPRAS paces budget by estimating future workload, and it assumes queries arrive in random order. Clustered bursts break that assumption, so our pacing uses a worst-case cap rather than a volume estimate. Accuracy-First Rényi DP shows that stopping rules that depend on the data can quietly break post-processing guarantees. That is why our trigger decisions are deliberately pure post-processing of an already-private scout output, and all sniper spending is charged at its worst case. Optimal Allocation of Privacy Budget on Hierarchical Data Release proves that static releases should give most of the budget to the finest level. It reports a tenfold bias and fourfold variance reduction over uniform splits on Wyoming census data. Its result is static. Our hypothesis is that in a stream, the budget should go to the finest level only where an alarm points.
Mechanism
Space is divided into a coarse grid, and each coarse cell contains $f$ fine cells. Time is discretized. The privacy unit is a person's contributions within any window of $w$ steps (w-event privacy), with each person's contributions clipped to one per step. The scout runs DP-CUSUM on each coarse cell's background-subtracted counts. The counts can go up or down relative to background, which the continual-release work cannot handle. Because coarse cells are disjoint, the scout costs $\epsilon_s$ per window across all cells. When a coarse cell alarms, the sniper releases Laplace-noised fine-cell counts over that cell and its neighbours for $L$ steps, at budget $\epsilon_k$. No more than $m$ sniper releases may affect any one person within a window. The guarantee is then
$$\epsilon_s + m\,\epsilon_k \le \epsilon,$$
where $\epsilon$ is the total per-window budget. The bound holds for any data, because both the alarms and the cap are computed from scout outputs that are already private. The design choices that remain are $f$, $L$, $m$, and the split between $\epsilon_s$ and $\epsilon_k$. We will set them using public simulated data, never the private stream. We will also analyze detection delay as a function of these settings by combining the DP-CUSUM delay bound for the scout with the localization error of an $L$-step Laplace release for the sniper.
Research Program
The work has three parts. The first is the mechanism and its privacy proof, including how to handle overlapping alarms from adjacent cells under the cap. The second is a delay and localization analysis that predicts when the two-tier design beats uniform allocation. Our working hypothesis is that it wins when burst duration $L$ is much shorter than $w$ and the burst occupies far fewer than $f$ fine cells, and loses otherwise. The third is an empirical study on two testbeds. The privacy-relevant testbed is the Gowalla check-in stream (6.4M check-ins), with synthetic hotspots injected so ground truth is known; the stream was previously used by LAPRAS and Budget Recycling DP. The stress-test testbed is a public Poisson sky simulator. It uses LEIA-like geometry (a 346 deg² field, 4–8 arcmin spot size, orbit-induced gaps), a SWGO-like 1:160 background ratio, and burst arrivals that cluster in time, such as repeaters and hour-long flares.
Baselines
The naive reference is static uniform allocation. It runs DP-CUSUM independently on every fine cell, with the per-window budget $\epsilon$ split evenly over time and parallel composition across cells. This is the strongest simple competitor because it pays no overhead for coarse aggregation and wastes nothing on the scout. It will use the same $\epsilon$, the same clipping, and threshold tuning on simulated data with an equal search budget.
The strongest existing method is the monitor-and-update scheme of Differentially Private Continual Release with Relative Error, adapted to fine-cell counts. It runs a noisy threshold test per cell and spends its update budget when that test fires. It is the closest competitor because it already concentrates budget at event times. What it lacks is the coarse-to-fine step. A second strong baseline is the static bottom-heavy hierarchical release of Optimal Allocation, applied at every step with DP-CUSUM on its leaves. We also include sliding-window ONLINEPCPD from Single and Multiple Change-Point Detection with Differential Privacy.
We found no reported result on Lacuna for DP-CUSUM or monitor-and-update on Gowalla-style spatial detection. DP-CUSUM's paper reports only qualitatively "significantly lower detection delays" than ONLINEPCPD on synthetic streams. The nearest reported numbers are on different tasks: LAPRAS reports up to 40% lower error than static splitting on Gowalla query answering, and the tenfold squared-bias reduction for hierarchical allocation is on Wyoming census data. We will therefore re-run every baseline ourselves.
Two ablations isolate the key ingredient. "Scout-only" gives all of $\epsilon$ to the scout and localizes from coarse output. "Sniper-without-scout" spends the same total budget on fine releases at random or scheduled times. Beating the baselines means at least a 25% reduction in median detection delay, or at least a 2× reduction in localization error, at matched false-alarm rate and matched $\epsilon$, across a range of $\epsilon$. It also means that neither ablation recovers that gain.
Why not just…?
Why not just split the budget uniformly per fine cell? For long, spread-out events this is probably as good as our method, and our analysis predicts that. Uniform allocation can only lose when events are short and local relative to the privacy window, which is the regime the scout-and-sniper structure targets. The experiments sweep burst duration and spatial extent so the crossover point is measured rather than asserted.
Why not just use monitor-and-update from continual release? This is the real threat, because it already spends budget only when a noisy test fires. The difference is that it spends at the resolution it monitors. To detect a small burst it must monitor fine cells, and testing many fine cells drives up false alarms. Watching coarse cells and switching to fine release on alarm is the one thing it does not do. The comparison at matched false-alarm rate, together with the scout-only ablation, settles whether that change in resolution is what produces any gain.
Why not just use event-level privacy, where each time step is independent and no allocation is needed? If one person's presence at a single time step is what we protect, budget allocation over time is unnecessary and this proposal has no purpose. We argue that location streams need person-level protection over time windows, because repeated check-ins identify people. If a reader rejects that threat model, the uniform baseline wins by construction.
Why not just pace the budget by predicted event volume, as LAPRAS does? LAPRAS assumes queries arrive in random order. Clustered transients violate that assumption, and its volume estimate would spend the budget too early during a cluster. We will include a LAPRAS-style volume pacer for the sniper budget on the clustered-arrival simulator to show this failure directly.
Decisive checks
Three outcomes would decide the question. The first is whether the two-tier method beats per-cell DP-CUSUM and monitor-and-update in delay and localization on injected Gowalla hotspots at $\epsilon \in \{0.5, 1, 2\}$ per window. The second is whether the measured crossover in burst duration and spatial extent matches the analytic prediction; if it does not, the delay model is wrong. The third is whether a membership-inference test on the released alarms and sniper counts stays within the stated budget, confirming that alarm-driven follow-up does not leak more than the cap allows. If the method wins only when bursts are extremely compact, we will report it as a niche tool rather than a general allocation rule.
References
- The Sub-TeV transient Gamma-Ray sky: challenges and opportunities
- The LOFT wide field monitor
- The Lobster Eye Imager for Astronomy Onboard the SATech-01 Satellite
- Sequential Change Detection With Differential Privacy
- Single and Multiple Change-Point Detection with Differential Privacy
- Differentially Private Continual Release with Relative Error
- LAPRAS : Learning-Augmented PRivate Answering for linear query Streams.
- Accuracy-First Rényi Differential Privacy and Post-Processing Immunity
- Optimal Allocation of Privacy Budget on Hierarchical Data Release
- Budget Recycling Differential Privacy