Planar Staircase Mechanism for AR Privacy

• Planar Staircase Mechanism (PSM) is a location privacy method that perturbs true positions using a piecewise-constant, staircase-shaped radial distribution while satisfying ε-geo-indistinguishability.
• PSM reduces expected displacement compared to the Planar Laplace Mechanism by concentrating more probability mass near the true location, thereby improving LB-AR quality of service.
• Its lightweight, constant-time design supports sub-millisecond execution for high-frequency updates and can be paired with thresholded reporting to mitigate trace-level privacy leakage.

Searching arXiv for the cited papers and closely related staircase-mechanism work. The Planar Staircase Mechanism (PSM) is a client-side location-privacy mechanism for real-time location-based augmented reality (LB-AR) systems in which user positions are modeled in the continuous plane $\mathbb{R}^2$. Introduced as a core component of PrivAR, PSM perturbs each true location by sampling from a staircase-shaped radial distribution that concentrates more probability mass near the true point than the Planar Laplace Mechanism (PLM), while retaining the same per-location $\epsilon$-geo-indistinguishability guarantee. In PrivAR, PSM is paired with Thresholded Reporting with PSM (TR-PSM) to address trajectory-level leakage under high-frequency reporting (Seeam et al., 4 Aug 2025).

1. Problem setting and design objective

PSM is motivated by the privacy requirements of LB-AR applications that stream sub-second GPS updates to preserve responsiveness and immersion. In this setting, the adversary is an honest-but-curious server that receives frequent location reports and may perform Bayesian inference on single points or exploit spatio-temporal correlations across a trace. The design target is therefore not only per-location privacy, but also low latency, low expected error, and preservation of quality of service (QoS) under continuous operation (Seeam et al., 4 Aug 2025).

The immediate baseline is PLM, which satisfies $\epsilon$-geo-indistinguishability by adding Euclidean noise. Its radial density is given as $f_R(r) = \epsilon^2 r e^{-\epsilon r}$. In the PrivAR formulation, this induces a probability peak at nonzero radius, specifically at $r = 1/\epsilon$, so the reported position is more likely to be displaced away from the true location. The consequence is increased expected error and degraded per-location QoS. A second limitation is compositional: when independent noise is added at every fix, a trace of length $T$ incurs privacy degradation to $T \cdot \epsilon$, which the PrivAR presentation describes as rapidly exhausting the privacy budget and making traces vulnerable to inference attacks. Other alternatives cited there—cryptographic LBS protocols, spatial cloaking, and classic LDP—are characterized respectively as too slow or heavy for client-side LB-AR, or as injecting too much noise or offering weak guarantees (Seeam et al., 4 Aug 2025).

Within this formulation, PSM is designed to be as lightweight as PLM while shifting probability mass toward the true location. The intended outcome is lower expected displacement for the same per-location privacy budget, without violating the real-time constraints of mobile AR workloads.

2. Construction and sampling rule

PSM perturbs a true location $x \in \mathbb{R}^2$ by sampling a radius $r \ge 0$ and an angle $\theta \in [0,2\pi)$, independently, and returning

$\epsilon$0

The angular component is uniform, $\epsilon$1, while the radial component is drawn from a staircase-shaped distribution rather than the smooth Laplace-type radial law used by PLM (Seeam et al., 4 Aug 2025).

The radial domain is partitioned into intervals of width $\epsilon$2, interpreted as steps, rings, or annuli. Within each interval the density is constant, and across intervals the mass decreases exponentially. The resulting distribution is piecewise constant in radius: probability is flat within a ring and drops only when the radius crosses a step boundary. This is the source of the term “staircase.” The parameter $\epsilon$3 denotes the number of steps, and the presentation notes that $\epsilon$4 is usually large, often with $\epsilon$5, so that the construction covers the plane (Seeam et al., 4 Aug 2025).

Parameter	Role
$\epsilon$6	GeoInd privacy budget
$\epsilon$7	Step width
$\epsilon$8	Number of steps

Sampling proceeds by inverse-CDF generation of the radial component. The algorithm given in PrivAR is: sample $\epsilon$9, compute $\epsilon$0 using the staircase radial CDF, sample $\epsilon$1, and output $\epsilon$2 (Seeam et al., 4 Aug 2025).

The construction is isotropic in angle but non-Laplacian in radius. That distinction is central: PSM is not merely PLM with discretized shells. Its utility advantage comes from the deliberate choice of a piecewise-constant radial profile with maximal mass in the innermost region.

3. Privacy guarantee and its scope

PSM is analyzed under $\epsilon$3-geo-indistinguishability, defined for a mechanism $\epsilon$4 by

$\epsilon$5

where $\epsilon$6 is Euclidean distance (Seeam et al., 4 Aug 2025).

The PrivAR analysis states that the worst-case privacy loss arises by comparing the maximum density in the annulus containing the true location to the density in a farther annulus indexed by $\epsilon$7. Because the step heights decrease exponentially, the density ratio is bounded by an exponential in the inter-point distance. Theorem 1 in that presentation is that PSM satisfies the same per-location $\epsilon$8-GeoInd guarantee as PLM (Seeam et al., 4 Aug 2025).

The privacy guarantee is therefore per release and per location. This scope is essential. PSM alone does not eliminate the composition problem for traces: repeated independent releases still incur cumulative loss. The PrivAR account explicitly states that, like PLM, repeated independent use causes cumulative budget loss for traces. A common misunderstanding is to treat the stronger utility of PSM as if it automatically implied stronger trajectory privacy. The mechanism does not make that claim; trajectory protection is introduced separately through thresholded reporting.

4. Accuracy, QoS, and computational profile

The main utility metric used for PSM is expected displacement, also expressed as Mean Normalized Error (MNE),

$\epsilon$9

For LB-AR end-to-end evaluation, the PrivAR study also uses application-level QoS measures such as the percentage of virtual objects reachable after perturbation and Gamescore (Seeam et al., 4 Aug 2025).

The reported empirical pattern is that PSM shifts the distributional mode from a nonzero radius to the origin. In the example stated for $f_R(r) = \epsilon^2 r e^{-\epsilon r}$0, PLM peaks at approximately $f_R(r) = \epsilon^2 r e^{-\epsilon r}$1, whereas PSM peaks at $f_R(r) = \epsilon^2 r e^{-\epsilon r}$2. Table I in the same description reports mean displacement of approximately $f_R(r) = \epsilon^2 r e^{-\epsilon r}$3 for PLM and approximately $f_R(r) = \epsilon^2 r e^{-\epsilon r}$4 for PSM at $f_R(r) = \epsilon^2 r e^{-\epsilon r}$5, described there as a $f_R(r) = \epsilon^2 r e^{-\epsilon r}$6 reduction. The summary results for PrivAR state that QoS (Gamescore) improves by up to 50%, attacker error increases by $f_R(r) = \epsilon^2 r e^{-\epsilon r}$7 over baseline, and the added runtime overhead is $f_R(r) = \epsilon^2 r e^{-\epsilon r}$8 milliseconds. The evaluation uses Geolife, T-drive, a custom GeoTrace AR dataset, and a Pokémon-Go-style prototype (Seeam et al., 4 Aug 2025).

The implementation profile is similarly emphasized. PSM and TR-PSM are described as constant-time primitives that can be executed in under $f_R(r) = \epsilon^2 r e^{-\epsilon r}$9 ms per fix, with negligible CPU and battery overhead on mobile clients. The significance is not merely asymptotic efficiency: the mechanism is intended for sub-second reporting loops, so the combination of low expected error and sub-millisecond execution is part of the design criterion rather than an incidental optimization (Seeam et al., 4 Aug 2025).

5. Thresholded Reporting with PSM

TR-PSM extends PSM from per-fix perturbation to selective reporting over a session. Its purpose is to reduce trace fragility by releasing a new noisy location only when the user’s displacement exceeds a privately randomized threshold. At session start, the mechanism samples a noisy threshold $r = 1/\epsilon$0, where $r = 1/\epsilon$1 is tunable and $r = 1/\epsilon$2 is sampled from the PSM radial noise distribution. For each new true location $r = 1/\epsilon$3, if $r = 1/\epsilon$4, the mechanism reuses the previous reported position; otherwise it emits a new PSM-perturbed location and updates the reference point (Seeam et al., 4 Aug 2025).

This creates a many-to-one mapping from true positions to reported positions. In the PrivAR formulation, that mapping is the main source of trace-level protection, because multiple consecutive real locations can collapse onto one released output. Theorem 2 states that each individual release remains $r = 1/\epsilon$5-GeoInd due to PSM, while the total session cost becomes $r = 1/\epsilon$6, where $r = 1/\epsilon$7 is the number of threshold crossings. This is contrasted with continuous reporting at cost $r = 1/\epsilon$8 (Seeam et al., 4 Aug 2025).

The mechanism also comes with explicit limitations. A large $r = 1/\epsilon$9 can increase perturbation and affect AR responsiveness. The threshold choice is empirical and context-dependent; it should exceed GPS jitter but remain small enough for interactivity. The PrivAR discussion also notes that some adversaries could exploit timing information, namely when updates are emitted and when they are not. These caveats delimit what TR-PSM addresses: it reduces trace inference risk under selective reporting, but does not erase every side channel associated with temporal interaction (Seeam et al., 4 Aug 2025).

6. Relation to staircase-mechanism theory and open questions

PSM belongs to a broader family of staircase mechanisms in differential privacy. The Podium mechanism introduces a finite mixture of three uniform distributions with truncated support and proves $T$0-DP. In the high-privacy regime $T$1, it is reported there to outperform Laplace and Staircase by 50–70% in variance reduction, while as $T$2 it asymptotically approaches the Staircase mechanism. That work is not a planar location mechanism, but it is relevant because it highlights the utility of piecewise-constant densities and the role of support geometry in variance control (Pihur, 2019).

A later result establishes a more general optimality theorem for vector-valued queries under additive $T$3-DP. Using convex rearrangement theory, that paper shows that the optimization over additive mechanisms reduces to radially symmetric distributions whose extreme points are staircase distributions, and proves that for any dimension, any norm, and any norm-monotone cost function, there exists an $T$4-DP staircase mechanism that is optimal among all additive mechanisms. The term “Planar Staircase Mechanism” does not appear explicitly there, but the theory specializes directly to $T$5 and characterizes radially symmetric staircase noise on annuli as the optimal 2D additive construction in that framework (Melbourne et al., 21 Jan 2026).

The relation between these results and PrivAR’s PSM requires care. The 2026 optimality theorem is stated for additive mechanisms under standard $T$6-DP with norm-based sensitivity and norm-monotone cost. PrivAR analyzes location perturbation under $T$7-geo-indistinguishability and evaluates end-to-end LB-AR QoS and trace privacy. This suggests a strong conceptual rationale for staircase-shaped planar noise, but it does not erase the narrower statement in the PrivAR presentation that PSM is not strictly proven optimal compared to linear-programming-based optimal geo-ind mechanisms. The two lines of work are therefore complementary rather than identical (Seeam et al., 4 Aug 2025).

Open directions are explicit in the PrivAR discussion. Repeated independent PSM releases still accumulate trace-level privacy loss; TR-PSM reduces but does not remove all leakage modes. Further refinements are suggested there through personalized thresholds, learned mobility models, and hybrid mechanisms that blend PSM with optimal map-based geo-ind mechanisms. A plausible implication is that future planar staircase designs will continue to balance three coupled constraints: local geometric privacy, temporal correlation control, and application-specific QoS under strict runtime budgets.