Regulators increasingly want to know which physical sources a regional environmental policy actually cut, such as coal combustion, steel dust, or traffic, and not only whether it lowered total pollution. The tool that answers this is positive matrix factorization (PMF), a receptor model that splits each measured air or soil sample into non-negative contributions from a small number of sources with non-negative chemical fingerprints. Monitoring data that could feed such an analysis at national scale often sits with firms, municipalities, and industrial parks, and these owners will not share raw concentrations. Differential privacy (DP), which adds calibrated noise so that no single participant's data measurably changes the output, is the obvious fix. Letting each participant choose its own privacy level is the obvious refinement. That refinement has a flaw. The chosen privacy level is itself informative. A heavy emitter in a resource-dependent city under carbon trading has the most reason to demand strong protection, so the demand for protection signals the emitter's status.
We propose to move privacy heterogeneity out of the individual and into the sampling design. Monitoring sites are drawn through a stratified multi-stage design whose strata are the policy contexts the evaluation needs anyway, for example resource-dependent versus non-resource cities or high versus low marketization. Each stratum is assigned one public privacy tier, and every site in it uses that tier. The privacy parameter then reveals nothing beyond stratum membership, which the published sampling frame already discloses. Because strata are known and their population weights are public, stratum-level estimates can be recombined without the bias that individual-level privacy weighting forces. The noise gap between strata becomes a sample-allocation problem: the design sends more sites to high-privacy strata to compensate.
Why this gap, and what already exists
Three lines of work motivate the proposal. Private Estimation When Data and Privacy Demands Are Correlated shows that when data values and privacy demands move together, weighting records by their privacy level beats a single conservative budget. Its heterogeneous frequency estimator reaches error 0.118 on UC Salary against 1.000 for the uniform baseline. The paper still assumes the analyst sees each person's privacy level. Managing Correlations in Data and Privacy Demand treats the pair (data, privacy choice) as the protected unit. Its Proposition 8 proves that no universal mechanism of this kind is unbiased across all correlation structures, because protecting conservative participants means down-weighting them. Sum Estimation under Personalized Local Differential Privacy and Context Aware Local Differential Privacy both flag hiding the privacy parameter as open. Proposition 8 is the constraint our design is built around. We do not claim an unbiased universal mechanism. We avoid needing one by making the correlation structure public and coarse, in the form of known strata.
On the factorization side, Decentralized Matrix Factorization with Heterogeneous Differential Privacy privatizes unconstrained matrix factorization for recommenders with per-rating privacy weights, using a "stretching" transform. It does not enforce non-negativity. It also leaves untested whether the server can infer a user's weight from the noise in the gradients that user sends. Heterogeneous Differential Privacy, where stretching originates, notes that it breaks for non-linear operators, and clipping to non-negative values is one of these. Privacy-Preserving Distributed Nonnegative Matrix Factorization does solve non-negative factorization across agents, using homomorphic encryption. It matches centralized error on MIT-CBCL faces within 40–60 iterations but gives no guarantee about the released factors. We reuse none of the stretching machinery. Instead, sites release noisy summary statistics, and the server projects onto non-negative solutions afterwards. Under DP, any computation on already-privatized outputs costs no additional privacy, so non-negativity is compatible with the guarantee.
On the application side, Chen et al. find carbon trading raised carbon efficiency more in resource-based cities (coefficient 0.0831) and in high-marketization cities (0.0845), with no significant effect in low-marketization cities. Xu et al. report a 10.9% average carbon reduction from China's environmental protection tax, with larger effects in non-resource and innovative cities. Both work with city-level aggregates, and neither can say which sources moved. The Yinchuan study Decadal transition of summertime PM2.5–O3 coupling and the Panzhihua campaign Source Characterization and Apportionment of PM10 in Panzhihua show that source composition depends on both place and policy phase. Yinchuan shows secondary organic aerosol rising to about 74% of organic mass. Panzhihua shows Ti-rich industrial fingerprints absent from standard libraries. Both analyses are centralized and use fixed receptor sites rather than a sampling frame.
Core Claim
In distributed PMF with per-site privacy levels, a curious server can recover each site's privacy level from the spread of that site's noisy updates, even when the level is never declared. The recovered levels predict policy-relevant site status. Assigning privacy tiers per design stratum removes this channel. Recombining stratum estimates with design weights removes the systematic under-representation of high-privacy strata that individual privacy weighting causes. Privacy-aware allocation of sites across strata recovers most of the variance that uniform-within-stratum noise costs.
Research Program
The data model has sites $s$, sampled within strata $h$. Each site holds a days-by-species matrix $X_s$ with measurement uncertainties, as in standard EPA PMF. All sites share one non-negative profile matrix $F$. Each site's contribution matrix $G_s$ stays local.
Fitting uses alternating non-negative least squares. In each round, a site computes its local $G_s$ exactly. It then releases the clipped, uncertainty-weighted products $G_s^\top W_s G_s$ and $G_s^\top W_s X_s$ with Gaussian noise at its stratum's tier $\epsilon_h$. The server sums these, solves for $F$ under non-negativity, and broadcasts it. The protected unit is a site's full series together with its tier assignment. We account for all rounds jointly using zero-concentrated DP composition.
For policy evaluation, each site also releases noisy stratum-period mean contributions per source. These enter a difference-in-differences regression on source-specific contributions. A difference-in-differences regression compares how treated and untreated units change before and after a policy. Its standard errors add the known noise variance to the sampling variance.
For allocation, suppose stratum $h$ has population weight $W_h$, between-site standard deviation $S_h$, $n_h$ sampled sites, and a released-mean noise constant $c$ determined by the clipping bound. The variance of the combined estimate is then approximately
$$V(n_1,\dots,n_H)=\sum_h W_h^2\left(\frac{S_h^2}{n_h}+\frac{c^2}{n_h^2\,\epsilon_h^2}\right).$$
The first term is ordinary stratified sampling variance. The second comes from averaging $n_h$ releases whose sensitivity shrinks as $1/n_h$. Minimizing $V$ under a fixed total or cost budget gives a privacy-aware version of Neyman allocation, the classical rule for splitting a sample across strata. We will derive it and compare it with proportional allocation.
The attack we will build is a noise-scale estimator. Across rounds, it compares each site's released statistics against the server's own prediction from the current $F$, then classifies the residual spread into tiers. A second stage maps the inferred tiers to site attributes.
Baselines
The naive reference point is a single uniform budget equal to the strictest demand, applied to every site. This baseline leaks nothing about tiers by construction. In related settings it has been close to useless: error 1.000 against 0.118 in Private Estimation When Data and Privacy Demands Are Correlated, and 85.77% relative error against 6.54% on MNIST in Sum Estimation under Personalized Local Differential Privacy. We still run it, because PMF's pooling over many days may behave differently.
The strongest competitor is individual-tier private PMF with the agnostic weighting of Chaudhuri and Courtade, where each site is weighted in proportion to $1-e^{-\epsilon_i}$. We use their add-remove accounting and the same summary-statistic protocol, so the only difference from our method is where heterogeneity lives. It is the strongest because it uses each site's own tolerance and wastes no privacy budget. No reported result was found on Lacuna for any private PMF or non-negative factorization on receptor data. The closest reported numbers are the 0.118 frequency-estimation error above and HDPMF's MSE reductions of 5.00% on ML-1M and 9.61% on Douban Decentralized Matrix Factorization with Heterogeneous Differential Privacy. Neither transfers directly. We therefore report our own runs.
We also adapt HDPMF's stretching with a projection step, to test Alaggan et al.'s warning that stretching fails under non-negativity. Encrypted distributed NMF as in PPDNMF and centralized non-private weighted PMF serve as accuracy ceilings, not privacy competitors.
All methods share the same simulated data, the same total privacy accounting at matched worst-case site guarantees, the same rank selection rule, and an equal hyperparameter search budget over clipping bounds and iteration counts.
Two ablations isolate our ingredients. The first uses stratum tiers with proportional allocation, which isolates the allocation rule. The second uses individual tiers with stratum-weighted recombination, which isolates the tier placement.
Beating the strongest competitor means three things at matched worst-case privacy. The tier-inference attack must drop to chance level for within-stratum attributes. The bias of the source-specific policy effect must fall to within its own confidence interval where the competitor's does not. The variance of that effect must be no more than modestly larger.
Why not just…?
Why not just use one conservative budget for everyone? It closes the leak completely, and it is the safest design if utility survives. The cited evidence from scalar and vector estimation says utility does not survive. The uniform arm will show whether that also holds for PMF, where pooling over days might soften the cost. If the uniform arm recovers profiles and policy effects nearly as well as ours, the tier machinery is unnecessary, and we will report that.
Why not just keep individual privacy levels but hide them from the server? Hiding the declared number does not hide the noise. The noise scale is visible in how much a site's releases scatter around the model's prediction across rounds, which is exactly the channel HDPMF left untested. Our first decisive check measures this directly. If the attack fails, individual tiers with the agnostic weighting become the better design, and the proposal narrows to the allocation result. This is the real threat, and the experiment settles it.
Why not just encrypt everything, as PPDNMF does? Encryption protects intermediate messages, but the released profiles and stratum contributions are the whole point of the analysis. Those outputs remain open to reconstruction and membership attacks, and PPDNMF gives no bound on them. We include it as an accuracy ceiling rather than as an alternative privacy guarantee.
Why not just average sites within each stratum and run ordinary PMF on the averages? Averaging first discards the day-to-day covariation across sites that lets PMF separate sources with overlapping fingerprints. That is the same weakness that sinks chemical mass balance under collinear profiles. We test it directly as a baseline on the Panzhihua-style scenario, where coal ash and steel dust overlap.
Proposed Experiments
No public multi-site, privacy-sensitive receptor dataset exists in the file, so the main testbed is semi-synthetic, and we say so plainly. Source profiles and shares are taken from published apportionments. The Panzhihua sources include coal combustion ash at 25.8%, iron and steel dust at 19.5% with Ti-rich markers, vehicle exhaust, and secondary sulfate. The Bangi study contributes five factors from EPA PMF 5.0 Fine particulate matter in the tropical environment.
Sites are nested in simulated cities. Cities are stratified by the resource-dependence and marketization contexts of the 282-city carbon-trading panel. A planted source-specific policy effect, for example a coal-factor reduction that is larger in resource cities, follows the direction reported by Chen et al. Privacy demand is correlated with each site's heavy-source contribution at controlled strength. Generic factorization sanity checks use the 30×200 synthetic matrix and MIT-CBCL faces from PPDNMF.
Metrics are chosen to follow the claim. Profile recovery is measured by cosine similarity after optimal factor matching. Stratum contribution error is measured against the planted values. The policy effect is scored on bias, confidence-interval coverage, and power. Leakage is measured as attacker AUC, first for inferring tiers and second for inferring heavy-emitter status within a stratum.
Three checks are decisive. The first tests whether the noise-scale attack recovers individual tiers, and through them emitter status, well above chance under individual-tier PMF, and whether it drops to chance under stratum tiers. The second tests whether, as the correlation between privacy demand and contribution grows, individual weighting shrinks the estimated resource-city coal effect while stratum recombination keeps it unbiased. The third tests whether privacy-aware allocation lowers policy-effect variance relative to proportional allocation at a fixed number of sites, in the direction the variance formula predicts.
If the first check fails, the leakage story collapses and the contribution reduces to design-based allocation. That outcome is itself a useful negative result for the heterogeneous-privacy literature.
References
- Private Estimation When Data and Privacy Demands Are Correlated
- Managing Correlations in Data and Privacy Demand
- Sum Estimation under Personalized Local Differential Privacy
- Context Aware Local Differential Privacy
- Decentralized Matrix Factorization with Heterogeneous Differential Privacy
- Heterogeneous Differential Privacy
- Privacy-Preserving Distributed Nonnegative Matrix Factorization
- Carbon Emission Trading Policy and Carbon Emission Efficiency: An Empirical Analysis of China’s Prefecture-Level Cities
- Identifying the Impact of Climate Policy on Urban Carbon Emissions: New Insights from China’s Environmental Protection Tax Reform
- Decadal transition of summertime PM 2.5 –O 3 coupling and secondary organic aerosol dominance in northwest China
- Source Characterization and Apportionment of PM10 in Panzhihua, China
- Fine particulate matter in the tropical environment: monsoonal effects, source apportionment, and health risk assessment