Federated learning promises privacy by shipping model updates instead of raw data, but the promise leaks: because a neural network is a differentiable function of its inputs, an adversary who intercepts a client's gradient can run a reverse optimization and reconstruct the private training image almost pixel-for-pixel. The original attack of this kind, Deep Leakage from Gradients, reaches a reconstruction mean squared error of about 0.0069 on CIFAR-100 where earlier shallow methods sat near 0.25 Deep Leakage from Gradients, and its stronger successor, which matches gradient direction rather than magnitude with a total-variation image prior, recovers roughly 14 dB PSNR images from a trained ResNet-20 where the original attack collapses to about 7 dB of unrecognizable noise Inverting Gradients - How easy is it to break privacy in federated learning?. Meanwhile, a parallel line of work in wireless communications has been pushing federated learning onto the radio spectrum using over-the-air computation: every client transmits its gradient simultaneously on the same frequency, and the physics of electromagnetic superposition adds the signals together in one channel use. That concurrency is exactly what wireless engineers call co-channel interference, and both communities treat it as either a communication trick or an engineering nuisance. This proposal argues it is neither: co-channel superposition is an unengineered privacy mechanism sitting inside the radio channel, and nobody has measured or designed it as one.
The core observation is simple. When sixteen clients transmit concurrently on one resource block, no individual client's gradient ever exists in isolation on the air — an eavesdropper can only capture a channel-weighted sum, entangled with fading and thermal noise. Two existing over-the-air federated learning systems exploit this superposition for speed but never evaluate it for privacy: the digital one-bit scheme with cooperative relays Cluster-Based Cooperative Digital Over-the-Air Aggregation for Wireless Federated Edge Learning and the probabilistic scheduler with unbiased reweighting Channel and Gradient-Importance Aware Device Scheduling for Over-the-Air Federated Learning. Both measure only convergence and test accuracy. Whether their co-channel transmissions resist modern gradient inversion is an open empirical question — and, we suspect, the answer is only partially yes, because inversion attacks with generative image priors are explicitly built to untangle batched, aggregate gradients Gradient Inversion with Generative Image Prior.
Core Claim
The claim has two parts, one negative and one constructive. First, the protection that co-channel superposition provides "for free" is real but shallower than assumed: it degrades Euclidean attacks badly, but direction-matching attacks with generative priors will still extract recognizable content from superposed updates at realistic cluster sizes and signal-to-noise ratios, and this has never been demonstrated. Second, the protection can be engineered to be much stronger at almost no cost, by adding interference on purpose. We propose filling spare capacity in the shared resource block with decoy-gradient transmitters: clients (or dedicated secondary devices) that synthesize concealed samples — synthetic images whose gradients mimic real ones but whose pixels look nothing like them, exactly the construction of DCS² Concealing Sensitive Samples against Gradient Leakage in Federated Learning — and transmit their decoy gradients over the same co-channel block. The base station receives real plus decoy gradients superposed and cannot separate them; an eavesdropper is in the same position plus extra channel noise. Because the decoys are gradient-mimetic, the aggregate remains a usable descent direction, so the privacy mixing rides along with the over-the-air aggregation at zero additional digital overhead — the channel does the mixing that DCS² currently does with local computation.
There is a pleasing inversion of wireless doctrine here. Classical physical-layer coexistence work projects secondary transmissions into the null space of the primary receiver so they cause no interference at all, as in the semi-unitary precoder built from the cross-channel null space for two-tier OFDM networks Cognitive interference alignment for OFDM two-tiered networks. Gradient privacy demands precisely the opposite: DCS² shows that inversion succeeds when sample gradients occupy orthogonal subspaces, so a defense must project decoys into the active gradient subspace to maximize entanglement Concealing Sensitive Samples against Gradient Leakage in Federated Learning. Our design is the first place where these two projection instincts meet: we deliberately violate null-space orthogonality over the air, but use power and phase control so the aggregate's first moment — the quantity the learning needs — stays correct.
Research Program
The program runs in three phases. Phase A is a systematic measurement study: we take the two published over-the-air systems — the one-bit majority-vote scheme with phase correction and greedy relay selection Cluster-Based Cooperative Digital Over-the-Air Aggregation for Wireless Federated Edge Learning and the probabilistic scheduler Channel and Gradient-Importance Aware Device Scheduling for Over-the-Air Federated Learning — and attack their transmissions with the full modern arsenal: Euclidean matching Deep Leakage from Gradients, cosine matching with a total-variation prior Inverting Gradients - How easy is it to break privacy in federated learning?, and generative-prior batch inversion Gradient Inversion with Generative Image Prior. We sweep cluster size, fading realization, and noise power to map where free co-channel protection breaks. This phase also covers the honest-but-curious-server case across multiple rounds, since sequential observation of updates is known to leak user-specific features beyond class averages Beyond Inferring Class Representatives: User-Level Privacy Leakage From Federated Learning.
Phase B is the constructive design. We formulate a joint scheduling and power-allocation problem: choose which clients and decoy transmitters share each resource block, and their transmit scales, to maximize attacker-side reconstruction error subject to two constraints — the base station's aggregate stays an unbiased estimator of the mean gradient within a variance budget (reusing the probabilistic-reweighting argument of the scheduler work Channel and Gradient-Importance Aware Device Scheduling for Over-the-Air Federated Learning), and any coexisting primary cellular tier keeps its quality of service. Formally, with $\mathbf{p}$ the transmit scales, $\mathcal{D}$ the scheduled decoy set, $\mathbf{y}_E$ the eavesdropper's received superposition and $\mathbf{y}_B$ the base station's:
$$\min_{\mathbf{p},\,\mathcal{D}} \;\mathrm{PSNR}_{\text{attacker}}(\mathbf{y}_E) \quad \text{s.t.} \quad \mathbb{E}\|\mathbf{y}_B - \bar{\mathbf{g}}\|^2 \le \tau,$$
where $\bar{\mathbf{g}}$ is the true mean gradient and $\tau$ a convergence-preserving budget. To make this tractable we borrow the convexification machinery proven for friendly-jamming resource allocation in heterogeneous networks, where non-convex interference constraints are converted to convex form using Perron–Frobenius spectral properties of the non-negative channel-gain matrices, converging in fewer than five iterations Secrecy-Optimized Resource Allocation for Device-to-Device Communication Underlaying Heterogeneous Networks. That paper established that scheduled secondary transmissions can act as "friendly jammers" against a wiretap eavesdropper; we are extending the same idea from Shannon secrecy rates to high-dimensional differentiable leakage, where the "secret" is a training image and the damage metric is reconstruction fidelity rather than bits.
Phase C is a theory component extending the information-theoretic friendly-jamming guarantees of Towards Secure Over-The-Air Computation, which bounds the eavesdropper's mean-squared estimation error from below for a scalar arithmetic average over idealized additive Gaussian noise channels. We extend the analysis to vector gradients under Rayleigh fading with exponential power delay profiles, following the multipath models used in the interference-alignment study Cognitive interference alignment for OFDM two-tiered networks. Crucially, we depart from that paper's architecture in one respect that matters: their jammer's signal is cancelled at the legitimate receiver, which hands an honest-but-curious receiver a clean update and therefore no protection at all against gradient inversion. Our decoy interference is deliberately not cancellable by anyone, including the base station; the guarantee we seek is not that the server learns nothing, but that what it learns is entangled with gradient-mimetic decoys whose visual content is dissimilar to any real sample.
Baselines
The do-nothing reference is undefended over-the-air transmission under the two published systems: the one-bit scheme run with 54 users in 6 clusters as reported, and the probabilistic scheduler with 30 non-IID devices at noise power $10^{-9}$ W Channel and Gradient-Importance Aware Device Scheduling for Over-the-Air Federated Learning, attacked by cosine matching with a total-variation prior, which is the strongest attack with a reported number on Lacuna: approximately 14 dB PSNR on a trained ResNet-20 over CIFAR-10, against roughly 7 dB for Euclidean matching Inverting Gradients - How easy is it to break privacy in federated learning?. This baseline tells us how much protection the channel gives for free; beating it means dropping reconstruction below recognizable quality (operationally, PSNR under about 10 dB and SSIM under 0.2) at matched test accuracy. The strongest existing defense is layer-wise mixed-precision quantization with locally retained scaling keys Mixed Quantization Enabled Federated Learning to Tackle Gradient Inversion Attacks, which on MNIST with LeNet-5 across 15 clients held 96.67% accuracy (versus roughly 78–85% for differential-privacy baselines) and drove Deep Leakage reconstruction PSNR from 55.3 down to 8.1; we run this defense digitally over the same simulated channel, the same datasets, the same client counts, and the same attack budget (identical inversion iterations and priors), so the comparison isolates where the privacy comes from rather than who tuned harder. We also run DCS² itself as a digital decoy baseline — reported numbers are reconstruction PSNR falling from 59.20 to 7.84 on MNIST and SSIM of 0.15 on CIFAR-10, with 65.68% test accuracy on TinyImageNet versus 52.06% for Soteria Concealing Sensitive Samples against Gradient Leakage in Federated Learning — because our decoy waveform is borrowed from it, and the fair question is whether transmitting decoys co-channel adds anything over mixing them digitally. Finally, two ablations isolate our key ingredient: our scheduler with decoys disabled (pure co-channel superposition and fading, no engineered interference), and the same decoys applied digitally with orthogonal, interference-free transmissions. If the full co-channel system does not beat both ablations, the "interference as privacy" ingredient is not earning its place and we will say so.
Why not just…?
Why not just add differential-privacy noise to the gradients? This is the default answer in the security literature, and the file contains consistent evidence that it pays heavily in utility. The original leakage paper found noise only stops reconstruction at variances above $10^{-2}$, at which point training accuracy collapses Deep Leakage from Gradients; the quantization paper reports differential-privacy baselines falling to roughly 78–85% accuracy where quantization holds 96.67% Mixed Quantization Enabled Federated Learning to Tackle Gradient Inversion Attacks; and per-example client DP, while accurate (97.48% versus a 97.98% non-private baseline on MNIST), pays a per-sample clipping and noising cost on every edge device Gradient-Leakage Resilient Federated Learning. Our scheme is not strictly better than DP — DP gives a provable guarantee we do not — but it targets a different operating point: privacy mixing that costs no added digital computation and no accuracy loss beyond the variance budget the channel already imposes.
Why not just quantize or prune the updates? This is the most serious alternative, and we treat it as the primary competitor rather than arguing it away. Mixed quantization is strong, cheap, and reported to defeat Deep Leakage outright Mixed Quantization Enabled Federated Learning to Tackle Gradient Inversion Attacks. Two things temper it. First, it has only been validated against the older Euclidean attack; the direction-matching attack recovers images from aggregated and pruned updates, and analytically proves the input to any fully connected layer is recoverable from that layer's weight gradients alone Inverting Gradients - How easy is it to break privacy in federated learning?. Second, and more honestly: it is possible that the one-bit quantized superposition at the heart of the digital AirComp scheme is already a good defense, and Phase A of our program is designed precisely to settle this — we attack the one-bit system with cosine matching plus a generative prior, which no paper in the file has done. If that combination fails to reconstruct, the constructive half of our proposal loses its motivation and the contribution collapses to the measurement study; we consider that a publishable but smaller outcome and accept the risk.
Why not just give each client its own frequency band so there is no interference at all? Orthogonal slicing eliminates co-channel interference but also eliminates the reason to share a channel, and the wireless literature is unambiguous about the cost. The foundational survey of device-to-device communication documents that strict orthogonal allocation severely under-utilizes spectrum A Survey on Device-to-Device Communication in Cellular Networks, graph-coloring work shows spatial reuse serves more concurrent links at higher sum rates than orthogonal configurations Joint Mode Selection and Resource Allocation for D2D Communications via Vertex Coloring, and both over-the-air learning papers emphasize that aggregation over a shared channel finishes in constant channel uses regardless of client count Cluster-Based Cooperative Digital Over-the-Air Aggregation for Wireless Federated Edge Learning. Orthogonal slicing also removes the entanglement that our defense exploits: separate bands mean individual gradients exist in isolation on the air, which is the best case for an eavesdropper.
Why not just use homomorphic encryption or secure multi-party computation? These give cryptographic guarantees that no physical-layer scheme can match, and we do not claim otherwise. The file's evidence is about feasibility: both the quantization defense and the secure-AirComp paper note that homomorphic encryption brings heavy computational latency and key-management overhead on edge devices Mixed Quantization Enabled Federated Learning to Tackle Gradient Inversion Attacks, Towards Secure Over-The-Air Computation, which is disqualifying for the low-power, real-time regime this proposal targets. Our claim is not superiority to cryptography but a near-zero-cost option for settings where cryptography cannot run — the same niche the finance and healthcare federated deployments occupy Federated learning enables big data for rare cancer boundary detection.
Proposed Experiments
Datasets and topologies follow the file's benchmarks so every number is comparable to prior work. MNIST with LeNet-5 across 15 clients matches the quantization study's setup, and the 54-user, 6-cluster topology matches the digital over-the-air scheme, whose cooperative relaying held final accuracy within 0.5% of the ideal majority vote Cluster-Based Cooperative Digital Over-the-Air Aggregation for Wireless Federated Edge Learning. CIFAR-10 uses the non-IID split of 30 devices holding 2 classes each at $10^{-9}$ W noise from the scheduler study Channel and Gradient-Importance Aware Device Scheduling for Over-the-Air Federated Learning, attacked on a trained ResNet-20 where the reported reference points are 14 dB (cosine) and 7 dB (Euclidean) PSNR Inverting Gradients - How easy is it to break privacy in federated learning?. CIFAR-100 carries the attack-fidelity comparison against Deep Leakage's reported MSE of roughly 0.0069 Deep Leakage from Gradients, and face datasets (LFW, CelebA) probe the most privacy-sensitive regime, where per-example DP reduced reconstructions to unrecognizable static Gradient-Leakage Resilient Federated Learning; no numerical reconstruction metric is reported on Lacuna for the face datasets, so we report ours qualitatively and by PSNR we measure ourselves. The wireless environment is simulated with Rayleigh fading and exponential power delay profiles as in the interference-alignment study Cognitive interference alignment for OFDM two-tiered networks, inside a 1000 m heterogeneous-cell layout with a passive eavesdropper as in the friendly-jamming resource-allocation work Secrecy-Optimized Resource Allocation for Device-to-Device Communication Underlaying Heterogeneous Networks. Metrics are attacker reconstruction PSNR, SSIM, and MSE per client; test accuracy versus communication rounds; aggregate gradient mean-squared error against the scheduler's unbiasedness budget; and channel uses and transmit energy per round. The decisive checks are three. First, does cosine-plus-generative-prior inversion defeat the published one-bit and analog over-the-air schemes at realistic cluster sizes — establishing that free channel protection is insufficient? Second, does the decoy-entangled scheduler keep test accuracy within one percent of the corresponding undefended over-the-air system while pushing attacker PSNR below the recognition threshold, beating mixed quantization run under the identical channel and attack budget? Third, does the Perron–Frobenius-style allocation converge within the handful of iterations its secrecy-rate ancestor needs, so that the privacy scheduling is fast enough for per-round re-allocation under user mobility? A positive answer to all three would establish co-channel interference as a designed privacy resource rather than a defect to be suppressed.
References
- Deep Leakage from Gradients
- Inverting Gradients - How easy is it to break privacy in federated learning?
- Cluster-Based Cooperative Digital Over-the-Air Aggregation for Wireless Federated Edge Learning
- Channel and Gradient-Importance Aware Device Scheduling for Over-the-Air Federated Learning
- Gradient Inversion with Generative Image Prior
- Concealing Sensitive Samples against Gradient Leakage in Federated Learning
- Cognitive interference alignment for OFDM two-tiered networks
- Beyond Inferring Class Representatives: User-Level Privacy Leakage From Federated Learning
- Secrecy-Optimized Resource Allocation for Device-to-Device Communication Underlaying Heterogeneous Networks
- Towards Secure Over-The-Air Computation
- Mixed Quantization Enabled Federated Learning to Tackle Gradient Inversion Attacks
- Gradient-Leakage Resilient Federated Learning
- A Survey on Device-to-Device Communication in Cellular Networks
- Joint Mode Selection and Resource Allocation for D2D Communications via Vertex Coloring
- Federated learning enables big data for rare cancer boundary detection