Some private decentralized learning schemes add noise that is designed to cancel. Each device perturbs what it sends to its neighbors with noise shares that sum to zero, so every individual message is masked while the network-wide average is left untouched. This "zero-sum" trick is what makes schemes like Zip-DL nearly free in accuracy. It depends on every share arriving exactly as sent. When devices exchange updates by analog over-the-air transmission and the devices are moving (vehicles, drones), that stops being true. Each share is multiplied by a wireless channel gain that has drifted since it was last measured, and the shares no longer sum to zero. The leftover noise piles up in the model.
We propose to design the noise correlations themselves around what a channel predictor knows about this drift. The key observation is simple. Noise shares that pass through the same, nearly unchanged channel cancel exactly even if that channel is estimated wrongly. Shares that pass through different or fast-changing channels cancel only as well as the channel is predicted. A forecaster of future channel states can supply how strongly any two transmissions' gains are expected to agree, and with what uncertainty. With that, we can choose how noise is spread across neighbors and across training rounds. The goal is for the cancelling combinations to go through gains that stay coherent, while no curious neighbor ever observes a combination that cancels. In short, the cost of cancelling noise should be set by channel prediction error, and noise can be placed where that error is small.
Core Claim
Consider analog decentralized learning over time-varying channels. Channel inversion uses predicted gains $\hat h$, so each transmission from node $i$ to receiver $j$ in round $t$ arrives scaled by $g_{ij}^t = h_{ij}^t/\hat h_{ij}^t$. Suppose the noise components $\xi$ sum to zero over some set $S$ of (sender, round) pairs. The part that fails to cancel at the receiver is then $r = \sum_{(i,t)\in S} (g_{ij}^t - \bar g)\,\xi_{ij}^t$, where $\bar g$ is the mean gain over $S$. Its expected power is
$$\mathbb{E}[r^2] = \mathrm{tr}\big(C_g\,\Sigma_\xi\big),$$
where $C_g$ is the covariance of the gain deviations over $S$ and $\Sigma_\xi$ is the covariance of the noise over $S$. Two consequences follow. First, zero-sum noise across neighbors (spatial cancellation, as in Zip-DL) pays a utility cost roughly proportional to the per-link prediction error. Second, zero-sum noise across consecutive rounds on the same link (temporal cancellation) costs almost nothing inside the channel's coherence time, whatever the estimation error. The catch is that a neighbor who sees both rounds can add them and read the sum of the updates with no noise at all.
The mechanism we propose, which we call coherence-aware correlated noise, resolves this tension. It solves a small, short-horizon matrix-factorization problem: minimize $\mathrm{tr}(C_g\Sigma_\xi)$, with $C_g$ taken from the channel predictor, subject to two constraints. The noise must cancel in the network aggregate, and every adversary's observation set must retain enough noise to meet a pairwise privacy target. We re-solve this problem every few rounds as new channel forecasts arrive.
Relation to Prior Work
The method builds most directly on Unified Privacy Guarantees for Decentralized Learning via Matrix Factorization. That paper, MAFALDA-SGD, writes decentralized training as $O_A = AG + BZ$, with the topology and the adversary's view encoded in fixed matrices. It optimizes noise correlations offline over all $T$ rounds, at a cost scaling as $O(n^3T + n^2T^3)$ for $n$ nodes. We reuse its accounting language, the adversary-view mask and pairwise network differential privacy. We change two things. $B$ becomes random through the channel, and the optimization runs over a receding window of $W$ rounds within local neighborhoods of degree $k$. By the same form, a window costs on the order of $k^3W + k^2W^3$, instead of one global offline solve that cannot track a drifting channel.
Low-Cost Privacy-Preserving Decentralized Learning, Zip-DL, supplies the zero-sum construction and motivates the problem. It proves cancellation only for lossless, symmetric links, and it degrades once more than 25% of nodes drop out. Topology-aware Differential Privacy for Decentralized Image Classification reuses noise across the graph and Noiseless Privacy-Preserving Decentralized Learning replaces noise with routing, but both likewise assume ideal digital links.
The temporal half of our design borrows from Correlated Noise Provably Beats Independent Noise for Differentially Private Learning. In that work, noise that is anti-correlated across steps cancels through linear training dynamics, so error scales with the effective rather than the ambient dimension. That paper considers a single centralized learner with no channel. We ask when temporal anti-correlation remains safe once a neighbor can see several rounds.
The wireless regime comes from Decentralized Wireless Federated Learning with Differential Privacy (DWFL), which aggregates over the air with power alignment. It explicitly leaves fast-fading, mobile channels as future work, because alignment computed from stale channel measurements fails. Adversary-Aware Private Inference over Wireless Channels shows that fading can itself act as privacy noise, but it too assumes block fading and a single uplink.
On the prediction side, MambaCSP forecasts fast-varying channels with a state-space sequence model at speeds up to 100 km/h. It has no privacy component. We use it, together with a classical autoregressive Kalman predictor, as the source of both $\hat h$ and $C_g$. Joint Optimization of Power Allocation and Training Duration for Uplink Multiuser MIMO Communications constrains the design. It shows that re-measuring the channel more often trades directly against data airtime, so we cannot simply sound the channel before every transmission.
Research Program
The first component is analysis. We will derive the residual-power expression above for AirComp-style decentralized SGD in which each receiver superposes its neighbors' signals. We will then bound how the residual propagates through the gossip averaging matrix over training, and give a convergence rate that depends on $\mathrm{tr}(C_g\Sigma_\xi)$ rather than on the full noise variance.
Privacy accounting has to survive the channel. We will assume conservatively that an adversary knows its own true channel, so fading gives no privacy credit. Because superposition weights depend on the realized gains, we will certify pairwise privacy uniformly over the predictor's confidence set of gains. We will also check how the guarantee degrades when the realized channel falls outside that set.
The second component is the mechanism itself. It is a per-neighborhood semidefinite program over the $kW$-dimensional noise covariance, with an aggregate-cancellation constraint and one privacy constraint per adversary view. Where possible we will derive closed-form solutions for two special cases, independent link errors and Jakes-type temporal correlation, to cut solve time.
The third component is a controlled study of how prediction quality shapes the achievable trade-off between privacy and accuracy. We will sweep predictor quality through user velocity, predictor class, and forecast horizon.
Baselines
The naive reference is DWFL-style independent Gaussian noise at each node, with channel inversion using the most recent measured channel. It has no cancellation to lose, so it is the fair "do nothing clever" point. Mobility hurts it only through misaligned superposition. We run it with the same privacy target, the same airtime (pilots plus data), and the same learning-rate tuning grid as every other method.
The strongest existing method on the privacy side is Zip-DL. We run it in two variants. The first uses stale channel measurements. The second, which is the real competitor, uses a plug-in MambaCSP or Kalman forecast of $\hat h$ but keeps Zip-DL's noise design unchanged. This plug-in variant isolates whether prediction alone suffices. We also run MAFALDA-SGD with its correlations optimized as if the channel were ideal, with a matched offline compute budget, to test whether a static optimal factorization breaks under aging. An oracle with perfect instantaneous channel knowledge gives the upper bound.
For reference numbers from the file, all measured on ideal channels: MAFALDA-SGD reports a 31% average improvement in test loss over DP-SGD on California Housing. Zip-DL reports up to 59% higher CIFAR-10 accuracy than Muffliato at equal membership-inference protection. MambaCSP reports 9% (TDD) and 12% (FDD) lower NMSE than LLM predictors on QuaDRiGa. No reported result was found on Lacuna for private decentralized learning under mobile, time-varying channels, so these baselines must be run fresh. Beating the plug-in Zip-DL means lower residual noise power and higher test accuracy at equal certified privacy, at velocities of 60 km/h and above, with the gap growing as velocity increases.
Three ablations isolate the key ingredient. The first uses only the spatial constraint. The second uses only temporal cancellation, and we expect it to show a measurable privacy leak under multi-round adversaries. The third replaces $C_g$ with a point estimate, which isolates whether the predictor's uncertainty matters and not just its mean.
Why not just…?
Why not just plug a good channel predictor into Zip-DL? This is the most serious threat. Under our analysis, the residual under plug-in prediction scales with per-link prediction error. If MambaCSP-level error is small enough at the tested speeds, the plug-in variant will essentially close the gap to the oracle and our mechanism is unnecessary. The velocity sweep settles this. Our claim is falsified if the plug-in variant stays within noise of our method across 10–100 km/h. We expect the difference to appear at high velocity and long feedback delay, where temporal coherence persists but point forecasts do not.
Why not just use independent Gaussian noise, which has nothing to cancel? Evidence from Correlated Noise Provably Beats Independent Noise for Differentially Private Learning and MAFALDA-SGD shows that independent noise scales with the ambient dimension, and that it diverges in strict-privacy regimes where correlated noise converges. It is nonetheless a plausible winner once aging destroys spatial cancellation. We report the velocity at which plug-in Zip-DL falls below this baseline, and whether our method stays above it.
Why not just transmit digitally with retransmissions, where noise shares arrive bit-exact? Digital links sidestep gain mismatch entirely. The remaining failure mode is packet loss, which Zip-DL tolerates up to about 25% dropout. What digital transmission gives up is the single-slot superposition that DWFL exploits. We therefore compare at equal total airtime, including retransmissions, and report accuracy per unit airtime. If digital transmission wins at our network sizes, the proposal's scope narrows to the dense, large-$N$ regime, and we will say so.
Proposed Experiments
Channels come from QuaDRiGa traces under 3GPP TR 38.901 NLOS, at 10–100 km/h, using the same configuration as MambaCSP. Traces are assigned to links in Erdős–Rényi, ring, and Facebook ego topologies. Learning tasks are FEMNIST, CIFAR-10 with ResNet-18, and California Housing.
The metrics are measured residual noise power against the predicted value $\mathrm{tr}(C_g\Sigma_\xi)$, test accuracy or loss, the certified pairwise privacy level, membership-inference AUC (Shatter reports a median of 58% as a near-random reference), and solve time per window.
We see two decisive checks. The first is that measured residual power tracks the predicted formula across velocities. The second is that at high mobility our method beats plug-in Zip-DL at equal certified privacy and equal airtime, while the temporal-only ablation visibly leaks under a multi-round adversary.
References
- Unified Privacy Guarantees for Decentralized Learning via Matrix Factorization
- Low-Cost Privacy-Preserving Decentralized Learning
- Topology-Aware Differential Privacy for Decentralized Image Classification
- Noiseless Privacy-Preserving Decentralized Learning
- Correlated Noise Provably Beats Independent Noise for Differentially Private Learning
- Decentralized Wireless Federated Learning With Differential Privacy
- Adversary-Aware Private Inference over Wireless Channels
- MambaCSP: Hybrid-Attention State Space Models for Hardware-Efficient Channel State Prediction
- Joint optimization of power allocation and training duration for uplink multiuser MIMO communications