When phones or sensors train a shared model by federated learning, they send model updates rather than raw data. Over-the-air aggregation lets all devices transmit their updates at the same moment as analog radio signals, so the channel itself adds the signals together. Radio noise then corrupts the sum. Earlier work showed that this noise can do useful work as differential privacy (DP), a formal guarantee that one person's data cannot be inferred from what is released. That work assumed a single hop from device to server. Devices that are far away need a relay. The cheapest kind, an amplify-and-forward (AF) relay, simply scales the analog signal it hears and retransmits it without decoding, so it amplifies its own receiver noise along with the signal. Relaying engineers treat that amplified noise as damage. We propose to count it as privacy, and to combine it with a shared random projection that compresses each update to $k \ll d$ numbers before transmission.
The central observation is simple. An AF relay multiplies the signal and its own first-hop noise by the same gain. The signal-to-noise ratio that sets privacy is therefore fixed at the relay's input, and nothing the relay or server does downstream can weaken it. That gives privacy against both a curious relay and a curious server by the post-processing property of DP. It also separates the privacy knob (device transmit scaling on the first hop) from the utility knob (relay gain and server scaling on the second hop). In single-hop systems these two are coupled through one receive-scaling factor. Physical noise is added per channel use, so projecting to $k$ dimensions changes the privacy level directly, not only the noise volume. We derive the resulting three-way trade-off between privacy, channel uses, and projection error, and test it end to end on non-i.i.d. federated benchmarks.
Core Claim
Consider $K$ devices behind one fixed-gain AF relay. Device $i$ clips its update to norm $C$, projects it with a seeded Gaussian matrix $P \in \mathbb{R}^{k\times d}$ shared by everyone, and sends $c\,P g_i$ over $k$ channel uses, after inverting only its own first-hop channel. The relay receives $y_1 = c\sum_i P g_i + n_1$ with noise variance $\sigma_1^2$ per use. It sends $G y_1$, and the server receives $G h_2 y_1 + n_2$. An average power limit $P_{\max}$ per channel use implies $c^2 C^2 \le k P_{\max}$. The per-round Gaussian-mechanism privacy loss seen by the relay, and hence by the server, then satisfies
$$\varepsilon_{\text{round}} \propto \frac{\Delta}{\sigma_1} \le \frac{\sqrt{k\,P_{\max}}}{\sigma_1},$$
where $\Delta = cC$ is the sensitivity measured at the relay's input.
This follows from the standard Gaussian mechanism with sensitivity $\Delta$ and noise standard deviation $\sigma_1$. Two consequences make the design new. First, at matched privacy, the noise-induced error in the server's average shrinks by about $\sqrt{d/k}$ compared with sending the full update at reduced power. The price is the projection distortion that random subspace training tolerates under DP. Second, the second-hop noise $n_2$ and any extra noise injected by the relay add privacy against the server only. This produces a two-tier guarantee: thermal noise $n_1$ protects devices from the relay, and $n_1$ plus $n_2$ plus relay top-up noise protects them from the server. Because devices invert only $h_1$, they never face the deep fades of the compound channel $h_1 G h_2$, which is far more likely to be near zero than a single Rayleigh link. That avoids the participation collapse that end-to-end inversion would cause.
Relation to Prior Work
Two papers motivate the privacy side. Leveraging the Physical Layer for Differential Privacy in Over-the-Air Federated Learning uses receiver thermal noise as free DP in a single-hop star and adds a dedicated cooperative jammer when the noise is not enough. CARES formalizes the trade-off governed by the server's receive scaling $\eta$ among noise, device truncation, and privacy, and it adapts $\eta$ with a bandit. Its envelopes are derived only for single-hop Rayleigh fading, and its 64-dimensional PCA is fixed offline. We differ on three points. Our topology is two-hop AF. The relay replaces the separate jammer. And the decoupling above removes the single-knob coupling that CARES has to manage.
Amplify-and-Forward Relaying for Hierarchical Over-the-Air Computation supplies the mechanism. It optimizes device, relay, and server coefficients to minimize the error of computing averages through AF relays, and the optimal solutions have a regularized channel-inversion form. It has no privacy, no learning, and no high-dimensional payloads. We reuse its decentralized partial-CSI design as the starting point and change its objective to minimize learning error under a privacy constraint. Hierarchical Over-the-Air Federated Edge Learning does train models hierarchically, but its intermediate servers decode digitally over an error-free backhaul and offer no DP. It therefore serves as our digital-relay comparison, not a precedent. Over-the-Air Inference over Multi-hop MIMO Networks shows that analog multi-hop chains can carry neural computation and spread wide vectors over time blocks. It performs inference only, with perfect CSI and no privacy.
The projection side comes from DP-GRAPE. It shows that data-independent Gaussian projections of gradients match SVD-based subspaces under DP, because DP noise flattens the singular spectrum. We use this to justify a seeded random $P$, which costs no privacy budget and needs no coordination beyond a shared seed. Efficient Private Empirical Risk Minimization for High-dimensional Learning provides the guarantee that noise then scales with $k$. PILLAR is the data-dependent alternative, using PCA computed on public data. Its strong gains, from 41.2% to 69.3% on Flower-16 at $\varepsilon=0.1$ with $k=40$, make it a necessary comparison for the projection choice. DP-GRAPE is centralized and injects software noise, and PILLAR trains linear heads on fixed features. Neither involves channels.
Research Program
The analysis has three parts. The first is a privacy accountant for the two-tier guarantee under fading. It uses zero-concentrated DP composition across rounds, tracks the realized per-round sensitivity at the relay after first-hop inversion and truncation, and adds relay top-up noise only when the thermal noise is not enough for the server-tier target. The second is a convergence bound for projected SGD under the effective noise $G^2|h_2|^2\sigma_1^2 + \sigma_2^2 + \sigma_{\text{top}}^2$ after server scaling. Combined with the privacy relation, it yields a predicted interior optimum $k^*$ for a fixed budget of channel uses per round. The third is a truncation analysis comparing split inversion with end-to-end inversion, extending the CARES envelope to the case where only $h_1$ is inverted.
The system we will build is a simulator. It contains Rayleigh and Nakagami fading on both hops, several relays each serving a cluster of devices, fixed-gain relays, and power limits on devices and relays. Real federated training runs on top of it, with seeded projections regenerated each round so that no subspace persists. The server lifts the aggregate back through $P^\top$ before taking the optimizer step.
Baselines
The naive reference is non-private FedAvg with ideal digital links, which gives the accuracy ceiling. Beside it we run AF over-the-air learning with no projection and no privacy accounting, which isolates the cost of the channel alone.
The strongest existing method on this question is CARES, given a two-hop channel. It is the only method in the file that jointly controls privacy, participation, and convergence for analog aggregation. We run it with end-to-end inversion over the compound channel, its bandit-tuned $\eta$, and its 64-dimensional PCA. Data, device count, privacy budget, channel uses per round, and tuning budget are identical to ours. CARES reports an improvement of 8.9 percentage points over fixed-scaling over-the-air baselines on FEMNIST with 100 clients and 62 classes, and 2.7 points on MNIST, in its single-hop setting. We will first reproduce its gap over fixed scaling in single hop, so that its implementation can be trusted, before moving it to two hops. Beating it means higher final FEMNIST accuracy at the same zCDP budget and the same number of channel uses, or equal accuracy with less device energy, consistently across five seeds.
The physical-layer DP design with a cooperative jammer is run over two hops. The jammer is placed at the relay's location and given the same total energy as our relay top-up. No reported accuracy result was found on Lacuna for this method on FEMNIST in a form we could match. We report our own reimplementation instead.
The digital comparison follows HOTAFL. Relays decode, add software DP noise, and forward over a wireless second hop using orthogonal slots. They are charged the channel uses this requires and the energy of a baseband processor. No reported accuracy number was found on Lacuna for HOTAFL.
Three ablations isolate the key ingredients. Full-dimensional transmission at power scaled down to the same $\varepsilon$ tests the $\sqrt{d/k}$ claim. Public-PCA projection, in the PILLAR style, replaces random $P$. End-to-end inversion replaces split inversion while everything else is held fixed.
Why not just…?
Why not just use digital decode-and-forward relays that add standard DP noise? This is the most serious threat when backhaul is cheap. A decoding relay sees the noisy superposition just as an AF relay does, so it gets no privacy advantage. Its costs are the extra orthogonal second-hop channel uses that grow with the number of relays, plus baseband processing at the relay. The low-power relay literature, for example Power Allocation and Cooperative Diversity in Two-Way Non-Regenerative Cognitive Radio Networks, cites these costs. Our comparison charges both channel uses and energy. If the digital design wins once those costs are counted, the AF case reduces to latency alone, and we will say so.
Why not just add noise at the relay, or stay single-hop with a jammer? Relay-side top-up is part of our design. What matters is that it protects against the server only. Against a curious relay, nothing downstream of $n_1$ helps. A single-hop jammer cannot be used where devices are out of range of the server, which is the regime that requires relays. The two-tier accountant makes this distinction measurable.
Why not just compress with parameter-efficient fine-tuning and send bits? Prompt tuning cuts payloads by about 90% (Federated Adaptive Prompt Tuning for Multi-Domain Collaborative Learning). Digital access latency still grows with the number of devices, and DP-GRAPE reports that private low-rank adapters underperform random gradient projection. The two approaches are complementary: projection can be applied to adapter updates. We test projected adapters against digital adapters at matched channel time.
Why not just use public-data PCA? When matched public data exists, PILLAR-style PCA may beat random projection. The ablation above decides this. DP-GRAPE's spectral-flattening result predicts that random projection will tie it at the privacy levels we target.
Proposed Experiments
Training uses FEMNIST from LEAF, with 100 writer-partitioned clients and 62 classes, and a 10-class non-i.i.d. variant in which each client sees five classes, following the physical-layer DP paper. We also train on CIFAR-10 with a small CNN under a Dirichlet split, and on MNIST for sanity checks. Privacy targets are $\varepsilon \in \{0.5, 1, 4.4\}$, matching operating points in the file. We sweep first-hop SNR, relay count, and $k$ from 16 up to $d$. Reported metrics are test accuracy, realized $\varepsilon$ in both tiers, device participation rate, total channel uses, and device energy.
Three checks decide the outcome. First, accuracy against $k$ at fixed channel uses must show the predicted interior optimum, and the full-dimensional ablation must lose by an amount consistent with $\sqrt{d/k}$. Second, split inversion must keep participation high in deep-fade regimes where end-to-end inversion collapses. Third, a privacy audit using membership inference with canary updates against the relay's received signal must not exceed the bound from our accountant. If the audit violates the bound, the accountant's handling of fading-dependent sensitivity is wrong, and the central claim fails.
References
- Leveraging the Physical Layer for Differential Privacy in Over-the-Air Federated Learning
- Certified Characterization of Privacy, Participation, and Convergence in Over-the-Air Federated Learning
- Amplify-and-Forward Relaying for Hierarchical Over-the-Air Computation
- Hierarchical Over-the-Air Federated Edge Learning
- Over-the-Air Inference over Multi-hop MIMO Networks
- Memory-Efficient Differentially Private Training with Gradient Random Projection
- Efficient Private Empirical Risk Minimization for High-dimensional Learning
- PILLAR: How to make semi-private learning more effective
- Power allocation and cooperative diversity in two-way non-regenerative cognitive radio networks
- Federated Adaptive Prompt Tuning for Multi-Domain Collaborative Learning
- LEAF: A Benchmark for Federated Settings