Drift-Safe Privacy from Quantum Hardware Noise

Recent work proposes using the noise that quantum computers already produce as a free privacy mechanism. Today's quantum processors, called NISQ devices (noisy intermediate-scale quantum: tens to hundreds of qubits, no error correction, high error rates), return random measurement outcomes ("shots") and corrupt those outcomes with gate errors. If that randomness is large enough, it can hide whether any one person's record was used, which is the guarantee of differential privacy, without adding the artificial noise that ruins utility on already-noisy hardware. The catch is that the guarantee is only as good as our knowledge of the noise.

Hardware noise drifts over hours. Shot noise alone also gives no worst-case protection, because a measurement whose outcome is nearly certain has almost no variance. The privacy claim therefore rests on an unobserved quantity, the device's current depolarizing rate, that is itself uncertain and moving. We propose to treat that rate as a quantity to be estimated with a confidence bound. The mechanism adds classical noise only for whatever privacy the lower confidence bound on hardware noise cannot cover. Downstream confidence intervals then use the matching upper bound on total variance. The result is a private release mechanism on NISQ hardware that stays private under drift (with an explicit failure probability folded into $\delta$) and gives honest error bars. Its noise cost should sit close to the "free privacy" ideal when the hardware is stable and fall back gracefully toward fully classical noise when it is not.

Core Claim

Harnessing Inherent Noises for Privacy Preservation in Quantum Machine Learning maps shot count and global depolarizing noise to an $(\epsilon,\delta)$ guarantee for parameter-shift gradients, using a central-limit argument. It applies probabilistic error cancellation (PEC, a technique that removes the average bias caused by gate errors at the price of extra variance) and reaches about 95% test accuracy on binary Iris. Toward a Hybrid Quantum Differential Privacy and Differential Privacy Preserving Distributed Quantum Computing similarly derive privacy budgets from depolarizing, damping, and measurement channels. All three take the channel parameters as fixed and known.

Adaptive Conformal Prediction for Quantum Machine Learning shows on ibm_sherbrooke that this assumption fails: noise drifts enough to break calibration. Those papers motivate the gap, and the drift result constrains any design that trusts a static noise model. Our claim is that a hardware-noise privacy guarantee is valid only once the rate is replaced by a high-probability lower bound estimated from data-independent probe circuits. That is the step none of these papers take.

The argument is short. Consider a one-shot $\pm1$ measurement of an observable whose noiseless mean is $\mu\in[-1,1]$, passed through global depolarizing noise of strength $p$. Its mean becomes $(1-p)\mu$, so its variance is at least $p(2-p)$. The $N$-shot average therefore has variance at least $p(2-p)/N$, whatever the data.

Let $\sigma^2_{\text{req}}(\epsilon,\delta,\Delta)$ be the Gaussian-mechanism variance needed at sensitivity $\Delta$. Let $\underline{p}$ be a $(1-\beta)$ lower confidence bound on $p$, estimated from probe circuits interleaved with the data circuits. We add classical Gaussian noise of variance

$$ \sigma^2_{\text{add}} = \max\Big(0,\; \sigma^2_{\text{req}} - \frac{\underline{p}(2-\underline{p})}{N}\Big), $$

which yields $(\epsilon, \delta+\beta)$ privacy whenever the noise is stationary within the interleaving window. Rescaling to undo depolarization is post-processing, so it does not change $\epsilon$. It does inflate variance, by roughly $1/(1-p)^2$ for global depolarizing noise. Confidence intervals on released quantities must therefore use the upper bound $\overline{p}$ together with the privacy variance.

This last step is the private uncertainty-quantification ingredient. Differentially Private Confidence Intervals for Empirical Risk Minimization and Differentially Private Ordinary Least Squares show that valid coverage requires modeling sampling and privacy noise jointly. Here the privacy noise's own variance is uncertain, a case those papers do not face.

Research Program

The first component is the noise-tracking estimator. We interleave short, data-independent characterization circuits, in the style of randomized benchmarking, with the data circuits. Because these probes never touch private records, they cost no privacy budget. We fit the effective depolarizing rate per window and form $\underline{p}$ and $\overline{p}$ by an exact binomial or bootstrap bound. We will test the stationarity-within-window assumption directly by comparing estimates from adjacent windows. We will also report the window length at which drift begins to show up as estimate disagreement.

The second component is the accounting. The Gaussian approximation in the Harnessing paper is asymptotic, so we will compare it against an exact accounting of the binomial-plus-Gaussian sum. Where the two disagree at small $N$, we use the conservative one. Across training steps we compose with standard Rényi accounting, and each step's $\beta$ enters the total $\delta$.

The third component is downstream uncertainty quantification on the same pipeline. One use case is private confidence intervals for aggregate observables computed over a sensitive dataset. These play the role of the sufficient statistics in Hypothesis Testing for Differentially Private Linear Regression, and we will use its parametric-bootstrap recipe with $\overline{p}$ plugged into the simulated noise. The other use case is a privately trained quantum classifier whose prediction sets are calibrated with the privatized quantile of Private Prediction Sets. That calibration runs on the shot-level nonconformity scores of Quantum Conformal Prediction for Reliable Uncertainty Quantification in Quantum Machine Learning, which supplies the scoring method. The calibration step is deliberately conventional. The new piece is that training and the error bars of the released statistics remain valid under drift, and that the training, calibration, and drift-adaptation budgets are accounted for together.

The trust model must be stated plainly. Hardware-noise privacy assumes the adversary does not see raw shots or control the device. We release only aggregated, post-processed values.

Baselines

The naive reference point is the static inherent-noise mechanism of Harnessing Inherent Noises for Privacy Preservation in Quantum Machine Learning. It sets $\epsilon$ from a single upfront estimate of the noise rate and adds nothing further. This is the baseline whose privacy claim we expect to break. The file reports about 95% binary-Iris test accuracy for it with PEC. We will rerun it under our drift schedules with identical circuits, shot budgets, and optimizer settings, and audit its realized privacy.

The strongest utility-safe competitor is fully classical noise injection in the style of DP-RMSProp from Quantum Machine Learning with Differential Privacy. That method ignores hardware noise, adds the whole $\sigma^2_{\text{req}}$ classically, and is private regardless of drift. It reports over 90% accuracy on binary MNIST for $\epsilon$ between 0.5 and 1.0, but in a noise-free simulator. On noisy backends it pays the double penalty, so we will run it on the same noisy backends and shot budget.

The key ablation is a plug-in variant that uses the point estimate of $p$ instead of $\underline{p}$. It isolates whether the confidence bound, rather than tracking alone, is what preserves privacy. For the conformal stage, the non-private references are QCP and AQCP. The file reports 90% target coverage for both, AQCP holding coverage with $\gamma=0.03$ on ibm_sherbrooke, HDR sets 1.5–2 times smaller than Euclidean ones at 1,000 shots, and up to 4.4 times smaller sets than standard CP for QCP in regression. These numbers bound how small private sets could possibly be.

Beating the baselines means three things. Our audited privacy must never exceed the claimed $\epsilon$ under drift, where the static mechanism does exceed it. Our accuracy and interval width must be measurably better than full classical injection at equal $\epsilon$. Our intervals must reach nominal coverage where the plug-in variant undercovers.

Why not just…?

Why not just add all the privacy noise classically and forget about hardware noise? This is the real threat. If the depolarizing contribution $p(2-p)/N$ is small next to $\sigma^2_{\text{req}}$ at practical shot counts, our method collapses to this baseline and its benefit is negligible. The Harnessing paper's report that added noise makes utility unusable on noisy hardware suggests the hardware share is substantial, but that is one dataset. Our experiments directly measure the fraction of $\sigma^2_{\text{req}}$ covered by $\underline{p}$ across shot counts and $\epsilon$ values. If that fraction is small, we will report that hardware-noise privacy is not worth its assumptions.

Why not just trust the vendor's published calibration data for the noise rate? Vendor calibrations are snapshots. The ibm_sherbrooke experiments in Adaptive Conformal Prediction for Quantum Machine Learning show that calibrations drift enough to break coverage within a session. The plug-in ablation, fed with snapshot rates, tests whether that drift also breaks privacy.

Why not just rely on shot noise, whose distribution is exactly binomial? Binomial variance $q(1-q)/N$ goes to zero when the outcome probability $q$ approaches 0 or 1. The worst-case privacy from shots alone is therefore nil. The lower bound on gate-induced depolarization is exactly what keeps outcomes away from certainty, which is why it must be estimated rather than assumed.

Why not just run AQCP-style online adaptation and privatize its updates? Online Differentially Private Conformal Prediction for Uncertainty Quantification shows how to privatize online coverage feedback. That keeps coverage under drift but says nothing about whether the training-time hardware privacy claim survives drift. We see it as complementary, and we include it as the drift-adaptation option in the conformal stage.

Proposed Experiments

Controlled drift comes from Qiskit's FakeQuitoV2 noise model, onto which we impose depolarizing-rate schedules: slow ramps down, steps, and oscillations matched in magnitude to the drift reported on ibm_sherbrooke. The tasks are binary Iris (as in the Harnessing paper), the 2D Moons and Concentric Circles sets and binary MNIST (as in the DP-QML paper), and the synthetic multimodal regression from AQCP for the conformal stage. A smaller replication on real IBM hardware checks that the estimator tracks real drift.

The metrics are audited privacy, model accuracy, interval coverage, and interval width. Privacy is audited empirically with canary records and the hypothesis-testing audit from Toward a Hybrid Quantum Differential Privacy. Accuracy is reported at matched $\epsilon$ and total shots. Coverage and width are measured for 95% intervals on released observables and for 90% prediction sets.

The decisive check is the downward-drift schedule. If the static mechanism's audited $\epsilon$ does not exceed its claim there, the core motivation fails and we will say so. If it does exceed its claim while ours holds, the second check is utility. Our mechanism should close most of the gap to the static mechanism's accuracy and remain clearly above full classical injection. A third check is coverage: intervals built with $\overline{p}$ should reach nominal coverage, and the plug-in variant should fall short under drift.