Does Free Channel Privacy Survive Unmixing Attacks? Leakage Maps for Over-the-Air Federated Learning

Wireless federated learning can get privacy "for free." When many phones send model updates at the same moment over the same radio band, the radio waves add up in the air. The base station receives only a noisy sum. Three lines of work turn that physical noise and mixing into a formal differential-privacy guarantee, which bounds how much any single training record can change what an observer sees. CARES tunes the receiver's scaling factor to certify privacy under fading. Uncoded transmission with adaptive power control proves a regime where channel noise alone meets the privacy target at no accuracy cost. Scaling-up analyses argue that adding more devices dilutes each device's information in the sum. None of these papers ever runs an attack. A separate literature has shown that summed updates are not safe. The Cocktail Party Attack treats a summed gradient as a mixture of signals and separates it back into individual private images, the way independent component analysis separates voices recorded by several microphones. It does this for 1,024 ImageNet images at once.

This note proposes to put the two together and measure, across the full operating range of an over-the-air system, how well state-of-the-art reconstruction attacks recover training images from the physically aggregated signal. It then tests a sharper prediction: the privacy credited to superposition and fading largely disappears against unmixing attacks, and what protects data is the additive noise per separable source. That quantity can be computed and can be used to choose radio settings. In two sentences: channel-based privacy proofs count the mixing of many devices' signals as protection, but separation attacks can undo mixing whenever the model's layers are wide enough. We will map where "free privacy" is real, where it is accounting without protection, and give a simple attack-calibrated rule for configuring the receiver.

Core Claim

In analog over-the-air aggregation, the receiver sees a signal of the form below. Here $\mathcal{S}_t$ is the set of devices whose fading coefficient $h_i$ cleared the truncation threshold in round $t$, $g_i$ is device $i$'s update, $n_t$ is thermal noise with per-coordinate variance $\sigma^2$, and $\eta_t$ is the receive scaling factor chosen by the server.

$$Y_t = \sum_{i \in \mathcal{S}_t} g_i + \frac{n_t}{\eta_t}$$

Our claim has three parts.

Mixing gives no protection while separation is possible. Within the regime where the fully connected layer has at least as many units as the total number of samples across participating devices (the condition the Cocktail Party Attack needs), adding devices or fading-induced gain mismatch adds no protection beyond what $\sigma^2/\eta_t^2$ alone provides. This is because unmixing is blind to how sources are scaled and combined.

A per-source signal-to-noise ratio should predict leakage better than the certified budget. Define it as each hidden sample's gradient energy in the fully connected layer divided by the noise energy $d\,\sigma^2/\eta_t^2$ over that layer's $d$ coordinates. It should predict reconstruction quality better than the certified privacy budget $\epsilon$ does.

There is a sharp collapse. The "blessing of scaling" becomes real protection only past the width threshold. There, reconstruction quality should drop abruptly rather than gradually.

If these hold, the existing channel-privacy certificates are correct as mathematics. But they mis-rank operating points by real risk, and the receiver should be configured with an attack-calibrated constraint.

Research Program

The first component is attack adaptation. We port three attack families to the received signal $Y_t$:

  • Cosine-angle inversion. We use the Inverting Gradients method, which matches gradient direction with an image-smoothness prior.
  • Generative-prior search. We use GGL and GIAS, which search a pretrained image generator's latent space and were designed to beat noise and sparsification defenses.
  • Noise-aware unmixing. We modify CPA so that the noise is whitened using the known $\eta_t$ before separation.

We assume an honest-but-curious server. It knows its own $\eta_t$, the channel estimates, and therefore $\mathcal{S}_t$, which is information it already needs to run the protocol. Attacks that ignore this knowledge would understate risk.

The second component is the leakage map. We sweep the following axes:

  • signal-to-noise ratio
  • $\eta_t$ (fixed values and the CARES bandit policy)
  • the number of devices $M$
  • local batch size
  • fully connected width
  • local steps (single-step SGD versus multi-step FedAvg, a gap CARES itself flags)
  • channel-estimation error

At every point we record test accuracy, the certified $\epsilon$ from the system's own accountant, and reconstruction quality.

The third component is the predictor and a simple fix. We regress reconstruction quality on $\epsilon$ and, separately, on the per-source signal-to-noise ratio together with the width-to-sources ratio, and compare how well each explains the data. If the per-source quantity wins, we add one constraint to CARES's feasible set of $\eta$: the predicted per-source ratio must stay below a threshold calibrated on held-out attack runs. We then check whether this buys lower leakage at equal accuracy.

Baselines

Undefended reference. Standard digital FedAvg with error-free transmission and no privacy mechanism. This anchors attack strength. The DCS² paper reports reconstruction PSNR of 59.20 with no defense on MNIST. Inverting Gradients reports about 14 dB on a trained ResNet-20 on CIFAR-10. Our attack implementations must reproduce numbers in this range on the undefended system before any defended result is trusted.

Strongest channel-native system. CARES is the most complete system: it couples noise, sensitivity and participation through $\eta$ and certifies the result under zero-concentrated differential privacy. It reports an 8.9 percentage-point accuracy gain over fixed-scaling over-the-air baselines on FEMNIST, and 2.7 points on MNIST compressed to 64 dimensions with principal component analysis. We reimplement it in exactly those settings, confirm the accuracy gains, and only then attack it. Fixed-$\eta$ over-the-air aggregation and the uncoded adaptive-power scheme serve as simpler channel baselines.

Matched synthetic central DP. This is error-free digital aggregation with Gaussian noise added to the sum, calibrated so its accountant reports the same $\epsilon$ as each over-the-air point. It isolates whether the channel differs from textbook noise at equal certified budget.

Digital secure aggregation. An exact sum with no noise, the setting in which CPA was demonstrated. Everything shares one data partition, model, number of rounds, and per-configuration attack budget measured in optimizer steps and GPU-hours. All defenses get the same tuning grid.

Ablations. These isolate the claim that superposition and fading add nothing against unmixing:

  • remove fading, keeping perfect alignment;
  • remove truncation, keeping full participation;
  • vary $M$ while holding the per-source signal-to-noise ratio fixed.

What counts as success. Our attack-calibrated CARES variant "beats" CARES if it lowers attack PSNR or SSIM at matched FEMNIST accuracy within one point. The diagnostic claim "beats" the $\epsilon$-based view if the per-source predictor explains clearly more variance in reconstruction quality across the map.

Why not just…?

Why not just reuse existing DP audits, since channel noise is Gaussian anyway? This is the most serious objection. At matched $\epsilon$, additive Gaussian noise at the receiver is distributionally the same as central DP noise, and GGL already showed that noise defenses can leak semantic content (PSNR above 5.7 dB). But those audits target single-client gradients with synthetic noise. They do not cover the extra structure of over-the-air systems:

  • participation that changes with $\eta$,
  • many-device superposition that scaling analyses credit as privacy,
  • the high-SNR operating points where "privacy for free" is claimed.

The matched-synthetic-DP baseline settles this directly. If the two are indistinguishable everywhere, the contribution narrows honestly to the scaling and width question and to the calibration rule. We would report that rather than claim a channel-specific effect.

Why not just tighten the privacy target until attacks fail? Deep Leakage from Gradients found that Gaussian noise stops inversion only at variances that badly hurt accuracy. Lowering $\epsilon$ in CARES also shrinks participation through the coupling of $\eta$. Our calibration constraint must beat "lower $\epsilon$ until PSNR is acceptable" at equal accuracy. If it does not, we will recommend the simpler rule.

Why not just use secure aggregation or a client-side decoy defense? CPA and fishing-model label attacks break exact sums, so secure aggregation is a weaker reference, not a fix. DCS² reaches 7.84 PSNR on MNIST and 65.68% TinyImageNet accuracy, but it needs per-sample decoy synthesis on device and gives no formal guarantee. We include it as a reference point on the accuracy–leakage plane, not as a competitor for battery-limited edge devices.

Proposed Experiments and Decisive Checks

Data and models. We use the CARES settings (MNIST compressed to 64 dimensions with principal component analysis, and FEMNIST with 100 clients) for exact replication, with reconstructions mapped back to pixels through the inverse projection. We then move to CIFAR-10 and TinyImageNet with convolutional networks whose fully connected width we control, using Rayleigh fading with configurable channel-estimation error.

Metrics. We measure:

  • test accuracy and worst-client accuracy;
  • certified $\epsilon$;
  • reconstruction PSNR and SSIM, plus label-recovery accuracy;
  • the fraction of samples recovered above a recognizability threshold, taken from the undefended runs.

Decisive checks. Four results decide the claims:

  1. If attack success at fixed per-source signal-to-noise ratio is flat in $M$ below the width threshold and collapses above it, the scaling claim is confirmed. A gradual decline instead would support the scaling-up view.
  2. If the per-source predictor ranks operating points by leakage while $\epsilon$ does not, certificates are mis-ranking real risk.
  3. If any configuration inside the published "privacy for free" region still yields recognizable reconstructions, that region is accounting-safe but empirically unsafe.
  4. If the matched-synthetic-DP baseline leaks the same as the channel at every point, the channel adds nothing beyond its noise, and we say so.