Over-the-Air Gradients Are Not Private

Federated learning promises privacy by having clients send only gradient updates — the derivatives of a loss with respect to model weights — instead of raw data. A newer wireless design, over-the-air computation (AirComp), goes further: all clients transmit their updates simultaneously in analog form, and the physics of the wireless channel sums them before any server ever sees an individual update. When this is combined with a cell-free architecture — many cheap, geographically spread access points (APs) connected by fiber links to one central processor — the approach fixes the classic "near-far" problem where far-away devices are drowned out, cutting aggregation error by nearly 10 dB versus a single-cell design Over-the-Air Computation in Cell-Free Massive MIMO Systems. This analog superposition is widely treated as a privacy feature: the central processor, and supposedly any eavesdropper, sees only a sum.

This proposal argues that assumption is wrong, and that the distributed antenna geometry making AirComp accurate is exactly what breaks its privacy. The reason is simple: power control that equalizes received signal strengths can only be equalizing on average across the AP network, but each individual AP sits close to some clients and far from others. Radio power falls off as roughly $d^{-3}$ in urban non-line-of-sight conditions 5G 3GPP-like Channel Models for Outdoor Urban Microcellular and Macrocellular Environments, so at any single AP the received analog sum is dominated by the handful of nearest clients — a local leakage hotspot that never appears in the clean aggregate at the central processor. An adversary who controls one access point, or taps one fronthaul link, receives a channel-weighted sum in which one client's gradient may be tens of dB stronger than the rest, and modern gradient-inversion attacks — which reconstruct training images from gradients — can recover that client's data.

Core Claim

The claim has three parts. First, the analog mixing in cell-free AirComp is not a privacy mechanism: at the level-2 cooperation tier defined by Chen, Björnson, and Fischione, where each AP computes a local sum and forwards a scalar to the central processor Over-the-Air Computation in Cell-Free Massive MIMO Systems, an honest-but-curious AP observes

$$y_l = \sum_{k=1}^{K} h_{lk}\, b_k\, g_k + n_l,$$

where $h_{lk}$ is the channel gain from client $k$ to AP $l$, $b_k$ is the client's transmit coefficient, $g_k$ its gradient, and $n_l$ receiver noise. Because $h_{lk}$ varies by orders of magnitude across $k$, the nearest clients dominate $y_l$, and their gradients can be isolated without any cryptanalysis. Second, even the central processor's vantage is weaker than assumed: across training rounds, the large-scale component of $h_{lk}$ is essentially fixed and partially estimable from public pilots, so the processor (or a fronthaul eavesdropper) can run a channel-aware disaggregation that separates client-scaled gradients from the round-by-round analog sums. A crucial technical point makes this practical: the strongest inversion losses are scale-invariant — Geiping et al. showed that matching gradient direction via cosine distance, not magnitude, is what makes inversion work on trained networks Inverting Gradients - How easy is it to break privacy in federated learning? — so the unknown per-client channel scale does not block reconstruction. Third, the defense should be co-designed with the physical layer rather than bolted on: the same transmit-coefficient freedom that Chen et al. optimize for aggregation accuracy can be randomized across rounds within an error budget, so that the effective mixing matrix is no longer recoverable and the disaggregation problem becomes ill-posed.

Research Program

The program builds a full adversarial evaluation of gradient privacy in cell-free over-the-air federated learning, then closes the vulnerabilities it finds. This fills a genuine gap: the existing bridge papers each stop short. The cell-free federated learning system of Federated Learning-Based Cell-Free Massive MIMO System for Privacy-Preserving derives a closed-form differential-privacy condition from low-resolution analog-to-digital converter noise and reports that 3-bit hardware performs nearly as well as 10-bit, but it runs zero empirical inversion attacks and assumes quantization noise behaves like ideal mathematical noise — an assumption that a diffusion-based denoising attack already falsifies in the digital setting, where Mjölnir achieves a PSNR of 18.25 on CIFAR-100 under $\epsilon = 1$ noise while the classical attack collapses to 0.112 Gradient Diffusion: A Perturbation-Resilient Gradient Leakage Attack. MOP-LOFPC Over-the-Air Federated Learning in Cell-Free MIMO with Long-term Power Constraint optimizes precoding and Lyapunov power control for convergence on a ridge-regression task but never tests whether its aggregated update can be inverted. Chen et al.'s AirComp analysis Over-the-Air Computation in Cell-Free Massive MIMO Systems contains no privacy analysis at all. And the one paper that breaks multi-client aggregation in the digital world, Gradient Disaggregation: Breaking Privacy in Federated Learning by Reconstructing the User Participant Matrix, assumes a binary participation matrix and error-free digital links; in AirComp the "matrix" is a continuum of complex channel gains, so its mixed-integer-programming machinery does not transfer directly, and we must build the analog analogue.

Concretely, we will build three attack surfaces of increasing sophistication and one defense. The AP-proximity attack needs no disaggregation: it models a compromised AP or fronthaul tap at level 2 and exploits channel-gain dominance, quantifying how many clients per AP are exposed as a function of AP density and path-loss exponent. The channel-aware disaggregation attack models the central processor or fronthaul eavesdropper across many rounds, estimating large-scale fading coefficients from pilot observations and solving a regularized bilinear recovery problem for client-scaled gradients, handling cross-round drift with the hard-threshold filtering idea from Gradient Disaggregation, then feeding the separated gradients into the strongest available inversion backends: Step-Wise GradInv, which reaches an LPIPS of 0.33 and PSNR of 15.12 at ImageNet batch size 8 and remains effective at batch size 64 High-Fidelity Gradient Inversion in Distributed Learning, and Mjölnir-style diffusion denoising Gradient Diffusion where perturbations are present. The denoising attack on "privacy for free" directly tests whether 1- and 3-bit quantization noise survives a generative prior, echoing how Gradient Inversion with Generative Image Prior constrains the under-determined search to the natural image manifold. The defense, privacy-aware transmit design, randomizes client transmit coefficients and AP combining across rounds subject to an aggregation-error constraint, reusing the alternating-optimization machinery of Over-the-Air Computation in Cell-Free Massive MIMO Systems and the causal power-control structure of Over-the-Air Federated Learning in Cell-Free MIMO with Long-term Power Constraint — trading a controlled amount of the cell-free MSE advantage for a mixing process the adversary cannot invert.

Baselines

The do-nothing reference point is the naive adversary that ignores the channel entirely: run Step-Wise GradInv directly on whatever sum a given vantage point receives, with the same tuning budget as our attacks. This is the right null model because it measures how much protection aggregation alone provides, and there is a calibration number for its digital cousin: Gradient Disaggregation reports that inverting the aggregated update on CIFAR yields a PSNR of only 6.1 — a failed reconstruction — and no reported result found on Lacuna for inversion of a channel-weighted over-the-air aggregate, so establishing this number is itself a contribution. We beat this baseline if channel-aware disaggregation or AP-proximity leakage materially raises reconstruction fidelity above the aggregate floor.

The strongest existing competitor is the gradient disaggregation attack of Gradient Disaggregation, which recovers the binary participation matrix with 100% accuracy for cohorts above 1,000 users and then lifts inversion PSNR from 6.1 on the aggregate to 36.5 on the separated updates, and sustains property-inference AUC above 0.9 where aggregated attacks fall to roughly 0.65. We will port it to the analog setting as the primary baseline — replacing its binary matrix with channel-estimate-based continuous weights — and give it the same telemetry, pilot observations, compute, and inversion backend as our attack, so the comparison isolates what the continuous wireless mixing adds or removes. Its digital result (PSNR 36.5) serves as the ceiling reference: our question is how much of that ceiling survives channel noise, fading, and power control. A second defense-side baseline is the quantization-based "privacy for free" scheme of Federated Learning-Based Cell-Free Massive MIMO System for Privacy-Preserving run at 1-, 3-, and 10-bit resolution under identical FL training, attacked with the same Mjölnir denoiser we use elsewhere; our defense beats it if it pushes image identifiability to chance level at a smaller cost in final model accuracy. Finally, an ablation reruns the full channel-aware attack with the estimated channel coefficients replaced by random guesses, which isolates whether leakage is genuinely driven by physical channel structure rather than by generic weakness of aggregation at large batch sizes.

Why not just…?

Why not just add differential-privacy noise to each client's update before transmission? This is the standard answer, and the evidence against it being sufficient is direct: Mjölnir models noise addition as a forward diffusion process and learns to reverse it, achieving a PSNR of 18.25 on CIFAR-100 at privacy budget $\epsilon = 1$ while the classical attack fails at 0.112 Gradient Diffusion: A Perturbation-Resilient Gradient Leakage Attack. Additive noise is also exactly the perturbation class that denoisers are best at removing, and it taxes convergence in a network where power is already scarce. We will still run DP noise as a baseline rather than argue around it, and report the privacy-accuracy curve honestly; if noise at a tolerable accuracy loss defeats all three of our attack surfaces, that is a legitimate finding and the proposal's co-designed defense would need to beat it to be worth deploying.

Why not just abandon AirComp and use digital secure aggregation? Cryptographic secure aggregation does not rely on physical assumptions, but it is not obviously safer: Gradient Disaggregation shows that even a server seeing only the exact sum $\sum_k g_k$ can recover individual updates across rounds using nothing but check-in metadata. It also forfeits the reasons AirComp exists: analog superposition removes the need for orthogonal uplink slots, and distributed remote units cut energy per bit nearly fourfold Radio-over-Fibre access for sustainable Digital Cities. The fair comparison, which our experiments include, is privacy per unit of communication resource — if secure aggregation plus DP dominates that frontier, the conclusion of the paper shifts from "fix AirComp" to "AirComp's privacy claim was never the reason to use it," which is still worth knowing. This is a real threat to the defense half of the proposal, and the Pareto-front experiment is designed to settle it rather than presuppose the answer.

Why not just run an off-the-shelf inversion attack on the aggregate and see if it fails? In the digital world it does fail — PSNR 6.1 on the aggregate Gradient Disaggregation — and if the same held at every vantage point in cell-free AirComp, the field's assumption would be vindicated and this proposal would be a negative-result paper. But two structural features make that outcome unlikely, and both are testable within the first experiment: the channel-weighted sum the AP sees is not the clean sum the digital server sees, and the AP-proximity hotspot is invisible from the central processor's vantage, so an aggregate-only evaluation would simply be measuring the wrong adversary. Our vantage-point-stratified design is what distinguishes this work from a rerun of existing attacks.

Why not just trust the closed-form differential-privacy guarantee from quantization noise? The guarantee in Federated Learning-Based Cell-Free Massive MIMO System for Privacy-Preserving is derived under an additive-noise model, and additive-noise DP is precisely what Mjölnir breaks at $\epsilon = 1$ with PSNR 18.25 Gradient Diffusion. The bound's authors also never ran an empirical attack, so the gap between the bound and achievable reconstruction is unmeasured. Our denoising experiment on 1- and 3-bit quantized systems directly tests whether the guarantee's premise survives contact with a learned denoiser; if the empirical leakage stays consistent with the bound, that would strengthen their claim, and we will say so.

Proposed Experiments

The evaluation stack is a cell-free AirComp federated learning simulator built on the 3GPP Urban Microcell geometry used by Chen et al., whose centralized level-3 design achieved aggregation MSE at least 9.7 dB lower than cellular massive MIMO Over-the-Air Computation in Cell-Free Massive MIMO Systems, with path-loss parameters drawn from the measured 3GPP models (urban macro non-line-of-sight exponent of 3.0) 5G 3GPP-like Channel Models for Outdoor Urban Microcellular and Macrocellular Environments. We train convolutional networks on CIFAR-10 and CIFAR-100 and an ImageNet subset — the settings where inversion quality is calibrated by prior numbers: PSNR of roughly 14 dB for direction-matching inversion of a trained ResNet-20 on CIFAR Inverting Gradients, LPIPS of 0.484 with about 28% of images identifiable at batch size 48 for GradInversion See Through Gradients: Image Batch Recovery via GradInversion, and LPIPS of 0.33 with PSNR of 15.12 at batch size 8 for Step-Wise GradInv High-Fidelity Gradient Inversion in Distributed Learning. Reconstruction fidelity is reported as PSNR, LPIPS, and image identifiability precision, plus property-inference AUC on an attribute-classification task in the style of the LFW experiments in Gradient Disaggregation; system cost is reported as AirComp MSE in dB, final test accuracy, and per-round energy, so that privacy and the communication gains that motivated AirComp are measured on one axis set.

The decisive checks are four. First, at the AP vantage, does the nearest-client hotspot let a single compromised AP reconstruct its dominant client's data at fidelity approaching the single-client reference of roughly 14 dB, and how quickly does that fidelity decay as AP density and client counts vary? Second, at the processor vantage, does channel-aware multi-round disaggregation lift inversion PSNR materially above the aggregate floor of 6.1, and how much of the gap toward the 36.5 digital ceiling survives fading and power control? Third, does transmit-coefficient randomization push image identifiability to chance while keeping level-2 aggregation error within a stated fraction of the 9.7 dB cell-free advantage — that is, does the privacy-accuracy Pareto front actually contain a usable operating point? Fourth, does the "privacy for free" quantization claim survive a Mjölnir-class denoiser at 1- and 3-bit resolution, or does learned denoising strip hardware noise as thoroughly as it strips injected DP noise? Each check has a pre-registered outcome in both directions, because the field currently assumes answers it has never measured, and either answer changes how over-the-air federated learning should be built.