Antenna Arrays Break the "Free Privacy" of Over-the-Air Federated Learning

In over-the-air federated learning, edge devices transmit their model updates simultaneously on the same frequency band, so the radio waves themselves add together and the base station receives the sum of all updates in a single time slot instead of one slot per device The Roadmap to 6G: AI Empowered Wireless Networks. A growing line of work argues that the thermal noise of the wireless channel is a gift: because the received sum is perturbed by natural Gaussian noise, it appears to satisfy differential privacy for free, without any device spending battery power broadcasting artificial noise Leveraging the Physical Layer for Differential Privacy in Over-the-Air Federated Learning; Strong Privacy and Utility Guarantee: Over-the-Air Statistical Estimation. The catch is that every one of these privacy guarantees is derived as if the server were a single antenna hearing one noisy number. Real base stations have antenna arrays, and an array does not hear one number — it hears a vector of differently-weighted copies of every device's transmission.

That observation breaks the free-privacy story in two distinct ways, and neither has been examined in the literature. First, when devices do not pre-correct their channel fading (as in channel-weighted or blind schemes), the server can invert the channel matrix and spatially separate individual clients before aggregation, reducing the problem to single-client gradient inversion — an attack setting where reconstruction of training images from gradients is devastatingly effective Inverting Gradients - How easy is it to break privacy in federated learning?; Deep Leakage from Gradients. Second, even when devices align their channels so the signals are inseparable, combining $K$ antennas shrinks the effective noise on the aggregate by a factor of $\sqrt{K}$, so the differential-privacy budget $\epsilon$ that was "paid for" by channel noise is quietly optimistic. This proposal audits both failure modes with real reconstruction attacks, then rebuilds the privacy-utility frontier for over-the-air learning under an adversary that owns the antenna array.

Core Claim

The claim is that the privacy guarantees of over-the-air federated learning, as currently derived, do not survive contact with a multi-antenna receiver, and that a redesign — rank-aligned precoding, low-dimensional updates, adversary-aware noise accounting, and scheduling more transmitters than antennas — restores protection using ambient channel noise alone, without the dedicated jammer hardware that current strict-privacy schemes require Leveraging the Physical Layer for Differential Privacy in Over-the-Air Federated Learning; Toward Secure and Private Over-the-Air Federated Learning.

The received signal at a server with $K$ antennas and $M$ participating devices is

$$\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{z}, \qquad \mathbf{H}\in\mathbb{C}^{K\times M},\ \mathbf{z}\sim\mathcal{CN}(0,\sigma^2\mathbf{I}),$$

where $\mathbf{x}$ stacks the devices' transmitted update vectors and column $i$ of $\mathbf{H}$ is device $i$'s channel to the array. For a Gaussian mechanism, the privacy budget scales as $\epsilon \propto C/\sigma$ with $C$ the per-client sensitivity bound. A server that combines antennas optimally faces effective noise of roughly $\sigma\sqrt{M/K}$ on each separated client when $K \ge M$ and $\mathbf{H}$ is well-conditioned, giving

$$\epsilon_{\text{true}} \;\approx\; \epsilon_{\text{claimed}}\sqrt{K/M},$$

and $\epsilon_{\text{true}} \approx \epsilon_{\text{claimed}}\sqrt{K}$ in the rank-one aligned case where clients cannot be separated but noise still averages across antennas. The scaling comes from standard Gaussian-mechanism algebra plus linear MMSE combining; the point is not the constant but the direction — more antennas mean less real privacy than the ledger says, and in the separable case the ledger is irrelevant because there is no aggregate to protect in the first place.

Research Program

The program has three parts: an attack audit, a corrected accounting theory, and a defense.

The attack audit runs the full modern suite of gradient-inversion attacks against the actual superimposed, faded, noisy waveform rather than a clean digital sum — something no paper in the literature has done. The attacks are: dummy-input gradient matching Deep Leakage from Gradients; cosine-similarity matching with image priors, which reconstructs images at roughly 14 dB PSNR from a trained ResNet-20 on CIFAR-10 versus about 7 dB for Euclidean matching Inverting Gradients - How easy is it to break privacy in federated learning?; Independent Component Analysis unmixing of the fully-connected-layer gradient, which breaks aggregated batches up to 1,024 samples but is known to degrade with additive noise and to fail when the layer is narrower than the number of mixed sources Cocktail Party Attack: Breaking Aggregation-Based Privacy in Federated Learning Using Independent Component Analysis; and the batch-norm "fishing" label attack, which recovers labels with 100% accuracy from exact digital sums Breaking Secure Aggregation: Label Leakage from Aggregated Gradients in Federated Learning. On top of these we add the new spatial disaggregation attack: zero-forcing or MMSE inversion of $\mathbf{H}$ at a multi-antenna server, followed by any single-client inversion method on the recovered per-device update. We will reuse the FILM gradient-inversion codebase FILM as an implementation scaffold alongside reimplementation of the ICA and fishing attacks. The audit varies the receiver architecture that the AirComp literature actually proposes — single-antenna aligned aggregation, channel-weighted post-processing where devices transmit at full power and the server reweights by channel gain Toward Secure and Private Over-the-Air Federated Learning, and blind transmission with massive arrays Blind Federated Edge Learning — because our hypothesis is that the utility-optimizing unaligned schemes are precisely the ones that hand the server a well-conditioned, invertible $\mathbf{H}$.

The theory part formalizes when disaggregation is impossible. Two conditions matter: the rank structure of the effective channel (clients that pre-invert toward a common beam produce a rank-one effective channel, which is unseparable by construction) and the noise-to-signal regime at which ICA unmixing becomes ill-conditioned. The Cocktail Party Attack paper acknowledges that additive noise hurts ICA but does not formalize the boundary Cocktail Party Attack: Breaking Aggregation-Based Privacy in Federated Learning Using Independent Component Analysis; we derive that boundary as a function of SNR, $K$, $M$, and network bottleneck width, and we recompute the Rényi-differential-privacy accounting of Private Wireless Federated Learning with Anonymous Over-the-Air Computation and the Moments-Accountant tracking of Leveraging the Physical Layer for Differential Privacy in Over-the-Air Federated Learning under worst-case antenna combining rather than single-antenna noise.

The defense combines four ingredients, each already motivated by a separate strand of prior work but never assembled. Devices pre-equalize toward a common server-broadcast beam so the effective channel stays rank-one and spatial demultiplexing is mathematically impossible, accepting the power cost of inversion. To recover the utility that alignment costs, and to shrink the noise burden, training is confined to low-rank adapter sub-networks — the perturbation-subspace-reduction result showing a 97.5% reduction in trainable parameters with minimal utility loss Managing the Privacy Utility Tradeoff in Deep Learning transfers here because fewer perturbed dimensions and smaller sensitivity let ambient channel noise cover the target $\epsilon$ even after the $\sqrt{K/M}$ erosion, which is what eliminates the cooperative jammer. Scheduling keeps $M > K$ participants per round so even a large-array server faces an under-determined inversion, amplified by client subsampling, which keeps privacy loss bounded over a thousand iterations at a 1% sampling rate Private Wireless Federated Learning with Anonymous Over-the-Air Computation. Finally, power control is tuned against the adversary-aware accounting, not the optimistic single-antenna ledger. Graceful degradation of analog aggregation — channel noise behaving like SGD stochasticity Machine Learning in the Air — is what gives us room to spend some SNR on privacy, and the COTAF precoding result showing $\mathcal{O}(1/T)$ convergence under heterogeneous data Over-the-Air Federated Learning From Heterogeneous Data tells us how much noise the training loop can absorb before convergence stalls.

Baselines

The naive reference point is aligned single-antenna over-the-air aggregation with no privacy mechanism beyond whatever channel noise happens to be present — the implicit default of the original AirComp designs Machine Learning in the Air. It sets both floors at once: it shows the convergence rate and test accuracy achievable with zero privacy spend, and it shows how badly undefended superimposed updates leak when our attacks are pointed at it. Beating it on privacy is trivial; the informative comparison is how little utility we pay relative to it.

The strongest existing defense is the physical-layer differential-privacy scheme of Leveraging the Physical Layer for Differential Privacy in Over-the-Air Federated Learning, which offloads DP noise onto a dedicated cooperative jammer and, at strict budgets $\epsilon \in \{1.0, 0.5, 0.1\}$ on non-i.i.d. FEMNIST with $\epsilon = 6.52, \delta = 10^{-5}$ and $\epsilon = 4.4, \delta = 0.01$ settings, reportedly beats the Liu & Simeone baseline in test accuracy and loss — though no reported result found on Lacuna for its absolute test-accuracy values. We will re-implement it faithfully and run it on identical data partitions, model architectures, SNR profiles, and tuning budget as our scheme, including its Upcycled-FL local-update schedule so the comparison isolates the privacy mechanism rather than the training cadence. Beating it means matching or exceeding its accuracy at $\epsilon = 1.0$ and $0.5$ without any jammer, and — the part it never tested — keeping reconstruction quality of our attacks low where its scheme, whose guarantee ignores multi-antenna combining, may fail once we point a $K$-antenna adversary at it.

The second competitor is the secure-and-private OTA-FL system of Toward Secure and Private Over-the-Air Federated Learning, whose channel-weighted post-processing is the most attractive utility design and therefore the most important to break: we run it on the same MNIST and CIFAR-10 setups at matched SNR and power budgets, and our attack audit will test whether its channel-reweighted sum exposes channel-favored clients to inversion, which its purely $(\epsilon,\delta)$ and MSE evaluation could not see. Its reported $\mathcal{O}(1/T)$ MNIST convergence is the utility bar.

The key ablation isolates the low-rank ingredient: our full defense versus the same rank-aligned, adversary-aware scheme with full-parameter updates plus a jammer standing in for the noise the adapters were supposed to make unnecessary. If the ablation matches the full scheme, the subspace story is wrong and the result reduces to "add more noise," which would be a genuinely useful negative finding. A second ablation drops only the $M > K$ scheduling, to separate the contribution of spatial crowding from that of rank alignment.

Why not just…?

Why not just run digital federated learning with standard DP-SGD? This is the safe option and we will include it as a utility reference at matched $\epsilon$. But it concedes the two advantages that motivate over-the-air aggregation at all: orthogonal collection scales as $\mathcal{O}(N)$ in latency and bandwidth The Roadmap to 6G: AI Empowered Wireless Networks, and local differential privacy forces every client to spend transmit power on broadcasting entropy rather than model signal, with minimax error decaying as $\mathcal{O}(1/n)$ against $\mathcal{O}(1/n^2)$ for over-the-air aggregation at the same privacy level Strong Privacy and Utility Guarantee: Over-the-Air Statistical Estimation. The honest caveat is that the $\mathcal{O}(1/n^2)$ result covers mean estimation, not deep learning; our experiments, which train real models under both regimes at matched $\epsilon$, will show whether the advantage survives outside that toy setting.

Why not just use cryptographic secure aggregation? Because the sum is not safe. ICA unmixing reconstructs images from aggregated batches up to 1,024 samples Cocktail Party Attack: Breaking Aggregation-Based Privacy in Federated Learning Using Independent Component Analysis, and a malicious server recovers individual clients' labels with 100% accuracy from securely aggregated sums Breaking Secure Aggregation: Label Leakage from Aggregated Gradients in Federated Learning. Secure aggregation also adds multi-round cryptographic overhead that sits poorly on bandwidth-starved edge radios. Our proposal should be read as studying the same vulnerability class those attacks exposed, in the channel regime where it has never been tested.

Why not just increase batch sizes or participants per round? Aggregation was once thought to confer privacy by averaging, and the attack literature has dismantled that: generative-prior inversion reconstructs images at batch size 32 with fidelity matching earlier attacks at batch size 4, at 15.58 dB PSNR versus 13.78 dB for pixel-space matching Gradient Inversion with Generative Image Prior, ICA breaks batches of 1,024 Cocktail Party Attack: Breaking Aggregation-Based Privacy in Federated Learning Using Independent Component Analysis, and individual images remain recognizable in batches of 100 Inverting Gradients - How easy is it to break privacy in federated learning?. Participant count still matters in our design — but as a scheduling constraint $M > K$ that makes spatial inversion under-determined, not as an averaging defense.

Why not just prune or quantize the updates before transmission? The measured thresholds are unattractive: pruning only blocks reconstruction above roughly 20% sparsity Deep Leakage from Gradients, while generative-prior attacks survive 99% sparsity Gradient Inversion with Generative Image Prior; half-precision floats do not help at all Deep Leakage from Gradients; and the label-fishing attack is resilient to gradient compression as long as utility is preserved Breaking Secure Aggregation: Label Leakage from Aggregated Gradients in Federated Learning. Notably, one-bit over-the-air aggregation One-Bit Over-the-Air Aggregation for Communication-Efficient Federated Edge Learning: Design and Convergence Analysis sits in this family, so our audit will include it as a defense-variant rather than dismiss it — if sign quantization genuinely neutralizes our attacks at acceptable accuracy, that is a finding we will report, not hide.

Proposed Experiments and Decisive Checks

All experiments run in a simulated physical layer with rayleigh-faded channels, per-device power constraints, and configurable $K$ and $M$, wrapped around standard federated training loops. Datasets and models follow the benchmarks in the file: MNIST with a CNN as in the secure-and-private OTA-FL study Toward Secure and Private Over-the-Air Federated Learning; CIFAR-10 with ResNet-20 for attack fidelity (the setting where cosine-similarity inversion reports ~14 dB PSNR on trained models Inverting Gradients - How easy is it to break privacy in federated learning?) and with a CNN for convergence under COTAF-style precoding Over-the-Air Federated Learning From Heterogeneous Data; non-i.i.d. FEMNIST restricted to ten heterogeneous classes exactly as in the physical-layer DP study, so the head-to-head with the jammer baseline is on its home turf Leveraging the Physical Layer for Differential Privacy in Over-the-Air Federated Learning; and Tiny-ImageNet with VGG-16 for the ICA attack at scale Cocktail Party Attack: Breaking Aggregation-Based Privacy in Federated Learning Using Independent Component Analysis.

The decisive checks are four. First, the spatial disaggregation attack on channel-weighted and blind schemes at $K = M$ should recover per-client updates well enough that downstream inversion approaches the clean single-client regime — PSNR in the neighborhood of the reported ~14 dB rather than collapsing; if it collapses instead, the free-privacy story is more robust than we claim and the paper becomes a defense of it. Second, on rank-aligned schemes, reconstruction fidelity at $K$ antennas should track the single-antenna case with noise reduced by $\sqrt{K}$, empirically confirming the $\epsilon$-erosion formula. Third, the defense at $\epsilon = 1.0$ and $0.5$ on FEMNIST should match or beat the jammer baseline's accuracy without any jammer, and the ablation without low-rank adapters should show a visible utility gap, isolating the subspace ingredient. Fourth, the theory's predicted ICA failure boundary — the SNR at which unmixing quality falls off — should match the empirically observed cliff on Tiny-ImageNet within a small factor. Privacy is scored with PSNR, LPIPS, and per-sample recognition, plus label-recovery accuracy for the fishing attack; utility is scored with test accuracy against training rounds and convergence rate. A negative result on any check is reportable as-is: the field currently believes channel noise buys privacy on the strength of a noise model nobody has attacked, and the first honest audit is valuable in either direction.