Free Privacy from Clock Drift: Policy-Graph Pseudo-Skew for Ad-Hoc Sensor Arrays

Every cheap sensor ticks at a slightly wrong rate. A phone microphone that should sample at 16,000 Hz might actually sample at 16,000.3 Hz, because its quartz crystal is imperfect, and that error is stable enough to identify the device. Cloaking the Clock shows that on car networks this "clock skew" works as a hardware fingerprint, and that an attacker who shifts message timing can fake one. Ad-hoc microphone arrays face the same problem. When a group of phones or smart speakers pool their recordings for beamforming, speech separation or self-localization, the server has to estimate each device's rate error, and so whoever runs it learns a fingerprint that links the same device across sessions.

This note proposes that each device resample its own stream, once per session, to a randomly chosen fake rate before upload. Downstream array processing already estimates and removes each device's total rate error, so this costs almost nothing in utility, while the published rate becomes provably hard to link to the device. The noise is calibrated with a policy graph. A policy graph comes from a generalization of differential privacy: you list which pairs of states must look alike to an observer, rather than protecting every pair equally. Here we only need to hide a device among other crystals of the same tolerance class. Taking the policy graph over the continuous skew axis, and using it on timing hardware rather than on databases or maps, is the new step.

Core Claim

Synchronization pipelines for distributed microphones are invariant to the identity of each device's clock. They need some consistent rate per device per session, not the true one. The mapping in Are Microphone Signals Alone Sufficient for TOA Self-Positioning? cancels unknown start times algebraically, with residual differences around $10^{-16}$ s. That means start offsets can be randomized uniformly over a large range at zero cost. Joint rate estimators such as Joint Optimization of Sampling Rate Offsets model a constant linear phase drift per device over a window. A constant injected skew stays inside that model, so it should also be nearly free, as long as the total skew stays within the estimator's working range.

Write device $i$'s published effective rate as $f_0(1 + s_i + z_i)$. Here $f_0$ is the nominal rate, $s_i$ is the true skew in parts per million (ppm), and $z_i$ is injected noise, drawn fresh each session. We discretize the skew range allowed by the manufacturer's tolerance into bins. We then place an edge between any two bins less than $\Delta$ ppm apart. This is the "geo-indistinguishability" special case of PGLP, but applied to a one-dimensional map of skews instead of a map of places.

Drawing $z_i$ from a Laplace distribution with scale $\Delta/\epsilon$ makes the published skews of any two adjacent bins $\epsilon$-indistinguishable. The bound from A Graph Symmetrization Bound on Channel Information Leakage Under Blowfish Privacy then turns this into an interpretable cap on how much an observer learns about the device's skew bin per session:

$$I_\infty \le \epsilon \cdot \mathrm{Diam}(G),$$

Here $I_\infty$ is min-entropy leakage, roughly the log-factor by which one observation raises the attacker's best guessing probability, and $\mathrm{Diam}(G)$ is the largest number of edges between two bins in a connected policy graph $G$.

That bound is proved only for finite domains, which is why we discretize. Extending it to the continuous skew line is an open theoretical piece of this work. The design choice that matters in practice is the trade-off between $\Delta$, the graph diameter and the noise scale on one side, and how far an estimator can track skew on the other.

Two limitations should be stated up front. First, fresh noise per session composes: $n$ linked sessions leak up to $n\epsilon$. We therefore treat the guarantee as per-epoch, analogous to MAC-address rotation, and measure how quickly real attackers actually accumulate evidence. Second, a constant pseudo-skew does not hide the shape of the crystal's temperature-drift curve, which may itself be a fingerprint. We measure that residual rather than assume it away.

Research Program

We first build the mechanism as an on-device step. It applies a random start offset and a random constant resampling factor, then uploads the result. The noise distribution and the policy graph are chosen per tolerance class.

Second, we characterize the utility cost as a function of injected skew magnitude. We expect it to be flat up to the range limit of the joint rate estimator and then to rise sharply. That curve sets the largest usable noise scale, and therefore the best achievable $\epsilon$ for a given $\Delta$.

Third, we build the attacker. This is a passive server that estimates each device's skew per session, and optionally ambient temperature, and tries to re-identify devices across sessions. It is the timing analogue of the trajectory-linking attack in User Identification across Asynchronous Mobility Trajectories, which matches users across disjoint time windows without perturbation.

Fourth, we study the analogue of PGLP's "isolated node" problem. An attacker who knows the room temperature and the crystal model can rule out some skew bins. Those rule-outs may cut a device's graph neighbors and quietly weaken the guarantee. We will test whether widening edges according to temperature uncertainty is enough, in the spirit of PGLP's graph repair.

We target passive observers of pooled sensor streams. We explicitly exclude safety-critical networks such as car buses, where skew-based intrusion detection is a protective feature, and the dual-use concern raised by Cloaking the Clock applies.

Baselines

The naive reference point is uploading raw streams. It gives the best utility and a clean fingerprint, and it sets both the utility ceiling and the attacker's best-case linking accuracy.

The strongest existing utility pipeline on real ad-hoc arrays is CTRnet from Cross-Talk Reduction. On CHiME-7 it reports a word error rate (DA-WER) of 22.6%, against 27.8% unprocessed and 26.6% for guided source separation. We will run CTRnet on CHiME-7 with and without our mechanism, using identical checkpoints and no retuning. Matching that 22.6% within run-to-run variance would count as "free" privacy.

For joint rate estimation we use the joint method of Joint Optimization of Sampling Rate Offsets, evaluated by rate-estimate error after removing the known injected skew. The file gives no reported number for this method, so we reproduce it ourselves on the same simulated 4-microphone setups.

For localization utility we use the offset-invariant self-positioning of Cao et al. on the Batstone office dataset and the 20-microphone simulated room. We also use the SLAM calibration of SLAM-based Joint Calibration of Multiple Asynchronous Microphone Arrays, which reports 0.233 m position error and drift estimation within 12.7 µs on its Turtlebot setup. Beating means staying within those errors under injected skew.

Two privacy baselines and one ablation complete the comparison. Standard differential privacy over the whole tolerance band is equivalent to a complete policy graph. Uniform random skew over that band adds no calibration at all. The ablation injects only start offsets and no skew. It isolates whether the skew noise, and not the free offset randomization, is what breaks linking. All variants share the same data, the same attacker, and the same tuning budget for the attacker's classifier.

Why not just…?

Why not just have each device correct itself to the nominal rate using network time or GPS? This is the most serious alternative. It removes most of the absolute skew without any noise. However, GPS is unavailable indoors, as the SLAM calibration work notes. Network time over Wi-Fi or Bluetooth suffers from the packet jitter documented in Elucidating Clock-drift. Most importantly, the residual after correction may still be a stable sub-ppm fingerprint. We will run self-correction as a baseline and measure the attacker's linking accuracy on the residual. If that residual turns out unlinkable, the idea reduces to "self-correct" and we will say so.

Why not just add huge uniform random skew and skip the graph? Larger skews push estimators out of their constant-drift operating range and accumulate misalignment faster. We expect the utility-versus-ppm curve to show this directly. The policy graph is justified only if small, $\Delta$-scaled noise buys most of the unlinkability, and the uniform-noise baseline tests exactly that.

Why not process everything on-device so raw timing never leaves? Beamforming and separation across devices need the pooled multichannel signal. Distributed models that avoid this, such as Attention-based distributed speech enhancement, assume perfect synchronization and do not address timing leakage at all.

Proposed Experiments

Simulated benchmarks give the utility side but not real crystals. We will therefore record a new fingerprinting corpus: about twenty phones of several models, many sessions each, recorded across temperatures, with simultaneous reference recordings. The file lists no public dataset of per-device skew identity, and no reported phone skew-fingerprinting accuracy was found on Lacuna. The closest number is the 100% IDS-deception success in Cloaking the Clock, which shows only that skew fingerprints are spoofable.

Three checks decide the question. The first is whether CHiME-7 word error rate, Batstone positioning error and rate-estimation error stay flat while the attacker's cross-session linking accuracy falls toward chance within an epoch. The second is how linking accuracy grows with the number of sessions, compared with the $n\epsilon$ composition bound. The third is whether temperature-curve shape lets the attacker recover identity despite constant pseudo-skew. If the third check succeeds, the mechanism must be extended to slowly varying injected skew, and its utility cost re-measured against the static-drift assumption of current estimators.