Privacy Certificates That Survive Fabrication Error in Analog Optical Aggregation

Some private federated learning systems get their privacy noise for free. Instead of having each client add random noise to its model update in software, they let the physical hardware supply it: thermal noise in a radio receiver, or photon shot noise in an optical detector. The privacy proofs for these "channel-native" schemes assume ideal hardware, where every client's signal reaches the aggregator with exactly the designed gain and the noise follows exactly the modeled distribution. Real analog hardware is never ideal. A coherent optical combiner fabricated with nanometer-scale lithographic errors routes each input with a slightly different strength. If client 7's beam splitter is 6% too transmissive, client 7's clipped update reaches the detector 6% larger than the proof assumes. That client's privacy loss is then larger than certified, and no software monitor will flag it.

This note proposes the first privacy accountant that includes fabrication variation. The accountant splits hardware deviation into two parts. The static, per-chip part can only inflate how much one client can move the output; it never adds privacy. The fluctuating, per-round part may be counted as privacy noise, but only under a certified lower bound on its variance. We pair the accountant with a statistical calibration procedure that issues per-device certificates, and with a circuit-topology analysis that picks combiner designs whose worst input port is least sensitive to fabrication error. The question we want to answer empirically is how large the silent privacy violation is under realistic fabrication tolerances, and how much utility a per-device certificate recovers compared with a blunt worst-case safety margin.

Core Claim

In analog aggregation, a client's privacy loss is set by the product of its clipping bound and the gain its signal experiences on the way to the detector. That gain is a fabrication-dependent random variable. For a Gaussian mechanism tracked with zero-concentrated differential privacy (zCDP, a privacy accounting scheme that adds up cleanly across training rounds), client $i$'s per-round privacy parameter is

$$\rho_i = \frac{(g_i C)^2}{2\,\sigma_{\min}^2},$$

where $C$ is the client-side clipping norm, $g_i$ is the realized amplitude gain from input port $i$ to the output after normalization, and $\sigma_{\min}$ is a certified lower bound on the standard deviation of the noise that is fresh in every round. The ideal-hardware analyses implicitly set $g_i = 1$ and set $\sigma_{\min}$ to the nominal noise level. Two consequences follow, and both are what make this work new.

First, a splitting-ratio error of a few percent inflates $\rho$ quadratically for whichever client lands on the strongest port. The certified budget for the whole system is set by that worst port.

Second, static fabrication disorder cannot be counted as privacy noise. It is drawn once per chip, so it repeats in every round, and an adversary who characterizes the device or averages across rounds removes it. The research file's own speculation (its Gap 2), that frozen lithographic disorder could act as a "zero-energy privacy shield," is therefore wrong as a differential-privacy claim. At best such disorder is a secret key, which is a different and weaker guarantee. This matches how Photonic Differential Privacy with Direct Feedback Alignment actually works. That paper runs training on an optical processor that multiplies by a fixed random scattering matrix, and its privacy proof draws only on shot noise and read noise, never on the frozen matrix. We formalize that distinction and extend it to fabrication-induced gain errors, which that paper does not model.

The certificate replaces $g_i$ with an upper confidence bound $\hat g_i^{+}$ obtained from calibration measurements. The result is a guarantee of the form $(\epsilon, \delta + \delta_{\text{cal}})$, where $\delta_{\text{cal}}$ is the probability that calibration underestimated the true gain.

Positioning Against Prior Work

The channel-native privacy literature supplies the mechanism we stress-test.

Certified Characterization of Privacy, Participation, and Convergence in Over-the-Air Federated Learning (the CARES method) adaptively chooses the receiver's scaling factor, which sets the trade-off between noise, privacy leakage and which clients can participate. It tracks leakage with zCDP. Leveraging the Physical Layer for Differential Privacy in Over-the-Air Federated Learning uses decentralized power control, plus a cooperative jammer node that injects interference when budgets are tight. Free Privacy Protection for Wireless Federated Learning maps bit error rates on digital links into Rényi differential privacy. Rethinking Federated Learning Over the Air: The Blessing of Scaling Up argues that adding clients strengthens privacy through superposition. All four assume that the client-side clipping bound is the sensitivity the server sees, which is exactly the assumption fabrication error breaks. The scaling-up argument is additionally constrained by unequal port gains, because superposition hides the strongest client least.

The fabrication literature motivates the threat and supplies our hardware model. Bayesian Photonic Accelerators measures static phase offsets and thermal crosstalk in meshes of Mach-Zehnder interferometers (MZIs, tunable two-path optical switches built from beam splitters). We reuse its distinction between passive per-chip offsets and drift, but apply it to privacy rather than energy. Effect of component variations on the gate fidelity in linear optical networks shows that two circuits computing the same transformation can differ greatly in fabrication sensitivity: one topology tolerated 3× larger errors on its outer beam splitters. We borrow that per-component sensitivity analysis to rank combiner topologies by their worst-port gain spread.

SEMU-Net pre-distorts layouts to cancel systematic over-etching and corner rounding. It also constrains us: it explicitly cannot remove stochastic shot-to-shot roughness, which is why some residual gain uncertainty must be certified rather than assumed away.

Research Program

We study a concrete system. Each of $K$ clients encodes its clipped update on the amplitude of an optical carrier. A silicon-photonic coherent combiner sums these signals; the candidate combiners are a binary tree of MZIs, a multimode interference coupler, and a programmable mesh. Photodetection at the server adds shot noise and read noise, which serve as the Gaussian privacy noise, as in Ohana et al.

Fabrication error enters as beam-splitter ratio errors and static phase offsets, drawn from distributions parameterized as in Sarantoglou et al. and Crickmore et al. Thermal drift enters as a slowly varying per-round phase term. Static phase errors only reduce the coherent gain through a cosine factor. Splitting-ratio errors can increase it, and those dominate the privacy risk.

The program has three parts.

Accounting theory. We derive the per-client bound above, including two complications. One is complex-valued gains, where phase error both shrinks a client's contribution and biases the aggregate. The other is the case where calibration-based pre-scaling has been applied, so the remaining uncertainty is in the calibration residual rather than the raw gain.

Calibration that yields certificates. The server injects known probe signals per port and estimates each $g_i$. Using the detector noise model, it forms a one-sided confidence bound $\hat g_i^{+}$ at level $\delta_{\text{cal}}$. Because thermal drift moves the gains, the calibration is refreshed on a schedule. Between refreshes, the certificate widens $\hat g_i^{+}$ by a drift envelope estimated from the drift model.

Topology selection. Using Crickmore-style per-component sensitivity maps, we compute for each combiner topology the distribution of $\max_i g_i$ under Monte Carlo fabrication draws. We then ask whether some topology concentrates its sensitivity into a few components that can be actively trimmed, leaving the rest passive.

For training we reuse the zCDP accounting and fixed-versus-adaptive scaling setup of CARES. That keeps the privacy bookkeeping comparable to the strongest channel-native method in the file.

Baselines

Naive reference: the ideal-hardware accountant. This is CARES-style zCDP accounting run on the fabrication-perturbed simulator, assuming $g_i = 1$. It is the baseline any deployed channel-native system would currently use. We compare it on the same training runs, clients and rounds. The comparison is not about utility, which is identical by construction. It is about whether its stated $\epsilon$ survives an empirical privacy audit. Beating it means showing that its stated $\epsilon$ is violated, with audit lower bounds exceeding the claim, while ours is not.

Global worst-case margin. This applies one safety factor to every client's sensitivity, using the 99.9th percentile of $\max_i g_i$ across Monte Carlo chip draws. It is the strongest simple competitor because it needs no per-device measurement. It is given the same noise level, clipping norm, and hyperparameter search budget as our method. Ours must achieve higher test accuracy at an equal certified $\epsilon$; we treat a gain under one point as a tie.

Strongest existing channel-native method: CARES with adaptive receive scaling. The file reports that CARES improves test accuracy by 8.9 percentage points over fixed-scaling over-the-air baselines on FEMNIST, and by 2.7 points on MNIST. No absolute accuracy is given in the file. We rerun CARES in our simulator with its guardrails corrected by our certificate, and report how much of its advantage persists once hardware variation is accounted for.

Hardware-noise privacy reference. The closest hardware-noise privacy result is Ohana et al.'s 83.36% FashionMNIST accuracy at noise level 0.05, against 86.80% non-private. We use it as a sanity anchor for the detector-noise regime, not as a head-to-head comparison, since that system is centralized. There is no reported result on Lacuna for private federated aggregation through a fabricated optical combiner.

Ablations. The first counts static offsets as privacy noise. This is the tempting error, and we expect audits to break it. The second uses per-device certificates without drift widening. The third holds the certificate fixed and swaps combiner topologies, which isolates the topology contribution.

Why not just…?

Why not just add a generous software noise margin and forget the hardware? This is the real threat. If the fabricated gain spread is only a couple of percent, a global margin costs little and our per-device machinery is overkill. Even then, choosing the margin requires a fabrication-aware bound on the worst-port gain, which is what our analysis supplies; without it, the margin is a guess. The global-margin baseline settles the question directly. If per-device certificates do not beat it by a meaningful amount at realistic tolerances, we will report that the contribution reduces to an accounting rule plus a recommended margin, and that is still a publishable correction.

Why not just calibrate the chip perfectly, using SEMU-Net-style pre-compensation or Bayesian phase priors? Calibration shrinks the gain error but cannot drive it to zero. SEMU-Net reaches 98.67% IoU on non-standard geometries and explicitly leaves stochastic roughness uncorrected. Thermal drift moves the gains after calibration, and Sarantoglou et al.'s power figures show that holding gains exactly by active tuning is expensive. Our ablation without drift widening tests whether a single calibration is enough.

Why not clip after the hardware instead of at the client? The gain is applied in the analog domain after the signals have already been superposed and detected together. The server never sees an individual client's contribution, so it cannot clip it. The only places to intervene are client-side pre-scaling, which is calibration, and the accounting itself.

Why not use digital secure aggregation and skip analog privacy entirely? That is a legitimate design point, but it forgoes the bandwidth and energy savings that motivate channel-native privacy. Li et al. also note that block ciphers amplify single bit errors over noisy links. This proposal asks how to make the analog route honest, not whether it beats cryptography.

Proposed Experiments

We train on three datasets. FEMNIST uses 100 non-IID clients, matching CARES. MNIST uses 64-dimensional PCA features, also matching CARES. FashionMNIST is included to compare against Ohana et al.'s noise regime. Fabrication tolerances are swept from 0 to 10% beam-splitter ratio error and from 0 to π/4 static phase error. Each operating point is evaluated over 200 or more simulated chips, with thermal drift injected on several timescales.

We report four metrics: certified $\epsilon$ at $\delta = 10^{-5}$, test accuracy, the per-client spread of privacy loss, and empirical $\epsilon$ lower bounds. The lower bounds come from a canary-insertion audit, in which a planted client update is distinguished from its absence across many runs.

Three checks are decisive.

The first is whether the ideal-hardware accountant's stated $\epsilon$ is exceeded by audit lower bounds at realistic tolerances, while our certificate holds in every chip draw. If the ideal accountant is never violated, the threat is not real and the project's premise fails.

The second is whether per-device certificates recover accuracy at a fixed certified $\epsilon$ compared with the global margin. We expect this to matter most where chip-to-chip variation is large compared with within-chip variation.

The third is whether combiner topology changes the worst-port gain distribution enough to shift the certified $\epsilon$ by more than the per-device certificate does. If so, fabrication sensitivity becomes a design criterion for private aggregators rather than just a nuisance to account for.