---
title: 'TR-PSM: Thresholded Reporting with PSM'
url: https://www.emergentmind.com/topics/thresholded-reporting-with-psm-tr-psm
type: topic
---

# TR-PSM: Thresholded Reporting with PSM

Searching arXiv for the primary TR-PSM paper and closely related works for disambiguation.
Thresholded Reporting with PSM (TR-PSM) is the selective release mechanism introduced within PrivAR for real-time privacy protection in location-based augmented reality (LB-AR). In this setting, PSM denotes the **Planar Staircase Mechanism**. PSM perturbs individual locations to provide per-location privacy, whereas TR-PSM adds a thresholded reporting policy: a noisy location update is released only when the displacement from the last released noisy location exceeds a private threshold. The resulting many-to-one mapping from multiple true locations to a repeated released point is intended to improve trace-level privacy while preserving the low latency and high quality of service required by sub-second GPS streaming in LB-AR applications [2508.02551].

## 1. Context and design objective

PrivAR studies privacy protection for LB-AR applications such as Pokémon-Go-style games and social AR platforms. The motivating setting is continuous, high-frequency, near-real-time location streaming from a trusted client device to an untrusted remote server, with a passive network observer also considered. The design target is threefold: **R1: low latency**, **R2: rigorous privacy**, both for each released point and for the whole trajectory, and **R3: high QoS**. The paper argues that conventional location-based services are less demanding because they issue relatively infrequent queries and can tolerate larger perturbation and higher latency, whereas LB-AR requires sub-second updates and interactive responsiveness [2508.02551].

The immediate technical problem is that independently perturbing every location exposes temporal structure. The paper identifies two specific limitations of repeated per-point protection. First, if a mechanism with per-location guarantee $\epsilon$-GeoInd is applied independently at every timestamp over a trajectory $\mathbf{x}=\{x_1,\dots,x_T\}$, the total privacy loss scales linearly as $T\epsilon$. Second, one-to-one release of perturbed points preserves strong spatio-temporal correlations, so an attacker can still exploit clustering around the underlying path. PrivAR addresses these two issues with a two-component architecture: PSM improves per-location utility relative to Planar Laplace, and TR-PSM reduces unnecessary releases so that trajectory privacy depends on threshold crossings rather than on every timestamp [2508.02551].

PSM itself is the base perturbation mechanism. It perturbs a true location $x\in\mathbb{R}^2$ by sampling a random radius $r$ and angle $\theta$, then outputting
$$
z=x+r(\cos\theta,\sin\theta), \qquad \theta\sim \mathrm{Unif}(0,2\pi).
$$
Its radial law is staircase-shaped rather than Laplace-shaped, and the paper states **Theorem 1**: PSM satisfies $\epsilon$-GeoInd for each location. TR-PSM does not replace this per-location mechanism; it uses it selectively [2508.02551].

## 2. Mechanism and algorithmic structure

TR-PSM is a session-based cached-state mechanism. Its inputs are a trace $\mathbf{x}=(x_1,\dots,x_T)$, a total budget $\epsilon_T$, a per-release budget $\epsilon$, and a nominal displacement threshold $\delta$. It maintains three pieces of state: a private threshold $\tilde{\delta}$, the most recently released noisy location $z_{\text{ref}}$, and a remaining budget counter $\epsilon_{\text{left}}$ [2508.02551].

Initialization consumes two units of privacy budget. First, threshold noise is sampled from PSM’s radial sampler,
$$
\eta \sim \mathcal{M}_r(\epsilon),
$$
and the private threshold is set to
$$
\tilde{\delta} \leftarrow \delta + \eta.
$$
Second, the first location is perturbed and released using PSM,
$$
z_1 \leftarrow \mathcal{M}(x_1,\epsilon), \qquad z_{\text{ref}} \leftarrow z_1,
$$
after which
$$
\epsilon_{\text{left}} \leftarrow \epsilon_T - 2\epsilon.
$$
If $\epsilon_T<2\epsilon$, the algorithm returns an error [2508.02551].

For each later timestamp $t\ge 2$, TR-PSM computes the Euclidean displacement from the current true location to the last released noisy location,
$$
d_t \leftarrow \|x_t - z_{\text{ref}}\|.
$$
The reporting rule is
$$
z_t =
\begin{cases}
z_{\text{ref}}, & d_t < \tilde{\delta},\\[4pt]
\mathcal{M}(x_t,\epsilon), & d_t \ge \tilde{\delta}\ \text{and budget available.}
\end{cases}
$$
When $d_t<\tilde{\delta}$, the previous report is reused rather than suppressed; the mechanism still outputs a stream element, but it is identical to the cached release. When $d_t\ge\tilde{\delta}$, a fresh PSM-perturbed point is released, $z_{\text{ref}}$ is updated to that point, and $\epsilon_{\text{left}}$ is reduced by $\epsilon$. If a threshold crossing occurs when $\epsilon_{\text{left}}<\epsilon$, the algorithm returns `BudgetExhausted` [2508.02551].

This structure is the essential source of the method’s many-to-one mapping. Instead of producing a one-to-one perturbed trajectory, TR-PSM may produce runs
$$
z_t=z_{t+1}=\cdots=z_{t'}
$$
even though the corresponding true locations differ. In PrivAR, that reuse policy is the defining distinction between PSM and TR-PSM [2508.02551].

## 3. Privacy model and formal guarantees

The privacy notion used is Geo-indistinguishability. PSM provides the per-location guarantee, and TR-PSM extends this to trajectory-level accounting by combining PSM releases with zero-cost reuse steps. The paper states **Theorem 2**: TR-PSM satisfies $\epsilon$-GeoInd for each location and $\epsilon_T$-GeoInd for a trace $\mathbf{x}$, where
$$
\epsilon_T=(k+2)\epsilon,
$$
and $k$ is the number of threshold crossings after the first fix [2508.02551].

The accounting is compositional. One unit of $\epsilon$ is spent to privatize the threshold, one unit is spent for the first PSM release, and each subsequent threshold crossing spends one additional $\epsilon$. Timestamps that merely reuse $z_{\text{ref}}$ incur $0$-GeoInd because no new information is released. By contrast, independent per-timestamp perturbation would spend
$$
T\epsilon
$$
over a trajectory of length $T$. TR-PSM therefore saves
$$
(T-k-2)\epsilon
$$
whenever $k+2<T$ [2508.02551].

The paper also gives a structural privacy argument beyond budget accounting. Independent perturbation yields a one-to-one sequence of noisy points, which still preserves temporal linkage. TR-PSM instead weakens sequential distinguishability because several nearby true locations map to the same released noisy point. In the paper’s phrasing, the mechanism creates many-to-one mappings that improve trace-level privacy [2508.02551].

The threat model assumes a trusted client device and an honest-but-curious server, with a passive eavesdropper also considered. The server is not trusted to preserve privacy and may monetize or leak location data. TR-PSM is implemented client-side, so the server receives only the privatized stream; no server-side logic changes are required [2508.02551].

## 4. Empirical behavior and operating trade-offs

PrivAR evaluates TR-PSM on two public datasets, Geolife and T-drive, and on the proprietary GeoTrace dataset, then validates it on a Pokémon-Go-style prototype. Geolife contains 1.05 million points with median step $8$ meters and $2$-second intervals. T-drive contains 809K points with median step $49$ meters and $181$-second intervals. GeoTrace was collected from five participants over about $127$ km across walking, running, biking, and driving [2508.02551].

The utility metric in the simulation study is Mean Normalized Error,
$$
\mathrm{MNE} = \frac{1}{|\mathcal{T}|} \sum_{j=1}^{|\mathcal{T}|} \frac{1}{T_j} \sum_{t=1}^{T_j} d(x_t^{(j)}, z_t^{(j)}),
$$
for which lower values are better. Privacy is evaluated with estimated Bayes risk, for which higher values indicate more attacker uncertainty. In the prototype, AR-specific QoS is measured by catchable objects and accumulated loss [2508.02551].

The principal empirical pattern is a three-way ranking. PSM has the best QoS, TR-PSM has slightly worse QoS than PSM, and TR-PSM still substantially outperforms Planar Laplace. In the trace privacy experiments at $\epsilon=0.1$, Bayes risk decreases for all mechanisms as trace length increases, but TR-PSM degrades slowest. The paper attributes this to many-to-one mappings and reused outputs that weaken sequential correlation. On public datasets, TR-PSM increases Bayes risk by up to $1.2\times$; on GeoTrace, it improves Bayes risk by up to $1.8\times$ over baseline. The abstract reports that PrivAR improves QoS (Gamescore) by up to $50\%$, while increasing attacker error by $1.8\times$ over baseline with an additional $0.06$ milliseconds runtime overhead [2508.02551].

The threshold parameter $\delta$ governs the central utility–privacy trade-off. If $\delta$ is too small, minor GPS jitter can trigger unnecessary releases. If $\delta$ is too large, updates are over-suppressed and stale-location error rises. The paper gives explicit MNE sensitivity values on Geolife: at $\epsilon=0.1$, MNE rises from $9.48$ at $\delta=5$ to $48.01$ at $\delta=100$; at $\epsilon=1$, it rises from $1.74$ at $\delta=5$ to $42.20$ at $\delta=100$ [2508.02551].

Prototype measurements place TR-PSM in the sub-millisecond regime. Mean runtime per location update is reported as $0.964$ ms on Galaxy A04, $0.152$ ms on Pixel 6a, and $0.069$ ms on Galaxy S22. In the end-to-end prototype, applying PrivAR costs only $0.05$–$0.07$ ms, less than $0.2\%$ of total end-to-end latency, which is dominated by network round-trip [2508.02551].

## 5. Terminological scope and adjacent literatures

The abbreviation **PSM** is overloaded across multiple research areas, and TR-PSM in PrivAR is specific to the **Planar Staircase Mechanism** [2508.02551]. This differs from several unrelated uses.

In causal inference and epidemiology, PSM commonly denotes **propensity score matching**. Methodological reviews of COVID-19-related studies identify omissions in covariate-selection justification, propensity-score model specification, balance diagnostics, and estimand reporting, while later methodological work argues that the so-called “PSM paradox” should not be treated as a reason to abandon matching [2403.07023] [2411.09579]. That literature concerns design and reporting in observational studies, not real-time location privacy.

In cryptography, PSM denotes **Private Simultaneous Messages**. The quadratic-residue line of work shows that threshold predicates are symmetric Boolean functions that can be realized in the PSM model with information-theoretic privacy, but this is a different primitive stack from PrivAR’s client-side noisy release mechanism [2209.01859]. Functionally adjacent telemetry systems such as STAR address threshold aggregation reporting through VOPRF, secret sharing, and symmetric encryption rather than through Planar Staircase noise and displacement-triggered reuse [2109.10074].

In optimization, PSM can denote the **parametric simplex method** for sparse learning [1704.01079]. In sleep staging, PSM can denote a **pressure-sensitive mat**, used as a sensing modality in multimodal fusion with EOG; that work is relevant to bed-based sensing but explicitly contains no thresholded reporting stage for PSM-derived signals [2506.06912]. The PrivAR meaning is therefore narrow and should be interpreted within the LB-AR privacy setting [2508.02551].

## 6. Limitations and open directions

TR-PSM’s main limitation is threshold sensitivity. Optimal $\delta$ values vary with mobility pattern and update frequency and must be empirically tuned. The paper states that TR-PSM is most beneficial on denser traces such as Geolife, where there are many closely spaced updates, while the advantage is smaller on sparser trajectories such as T-drive [2508.02551].

A second limitation is explicit session-level budget management. If a threshold crossing occurs when
$$
\epsilon_{\text{left}} < \epsilon,
$$
the mechanism returns `BudgetExhausted`. The paper therefore requires either a sufficiently large session budget or session restart logic at the application layer [2508.02551].

A third limitation is that TR-PSM trades some freshness for trace privacy. When the user moves quickly or when $\delta$ is large, reuse of $z_{\text{ref}}$ can produce stale reports that degrade AR quality. The results show that this degradation is typically modest relative to Planar Laplace, but it is intrinsic to the mechanism [2508.02551].

The paper does not present a learned or theoretically optimized thresholding policy. It explicitly suggests tuning $\delta$ based on application update rate and movement characteristics. This suggests that adaptive threshold selection remains open. A plausible implication is that future work could preserve the exact release rule of TR-PSM while modifying how $\delta$ or $\tilde{\delta}$ is chosen across sessions or mobility regimes, but that extension is not part of the reported method [2508.02551].

Source: https://www.emergentmind.com/topics/thresholded-reporting-with-psm-tr-psm