Some federated learning methods cut communication by having each client send a handful of synthetic training examples instead of its model update. Federated learning trains one shared model across many devices without collecting their data. The server then trains on those synthetic examples to recover the update. The trouble is how the examples are made. The client optimizes them until the gradient they produce matches the gradient of its real data, and that is exactly the optimization a gradient-inversion attacker runs to reconstruct private training data from intercepted gradients. A client that does this has, in effect, run the attack on itself and mailed the result to the server.
This note proposes to measure how much private data these synthetic-data uplinks expose, compared with ordinary compressed gradients at the same number of transmitted bits. It also proposes a modified synthesis objective that keeps the update-matching property while forcing the synthetic examples away from any individual private example. The opportunity is concrete. FedSynth explicitly leaves open whether its synthetic datasets leak private data, and FedDM reports roughly 80% per-round communication savings with a related scheme. If these methods hand over near-copies of client data, their efficiency comes at a cost nobody has accounted for.
Core Claim
We claim that leakage from a synthetic-data uplink is governed by the ratio of synthetic examples to real local examples, not by payload size. When a client holds few examples relative to the synthetic budget, the easiest way to match its gradient is to reproduce its data. Leakage should therefore be worst exactly where communication savings look best: small, data-poor clients such as those in FEMNIST.
We further claim the following comparison. Gradient-matching proxies (FedSynth) should leak individual samples. Distribution-matching proxies (FedDM), which match feature statistics rather than per-example gradients, should leak less per sample but still reveal which data were used, measurable by membership inference. Membership inference is an attack that decides whether a given example was in a client's training set.
The constructive claim is a repulsion-constrained synthesis. It adds a penalty that keeps every synthetic example at least a margin away from every real local example in a fixed feature space. The resulting point should sit at a better position on the three-way trade-off between bits sent, accuracy, and exposure than either unconstrained proxies or compressed gradients.
The objective is
$$\min_{S}\; \big\| g_\theta(S) - g_\theta(D_k) \big\|^2 \;+\; \lambda \sum_{s \in S} \max\!\Big(0,\; \tau - \min_{x \in D_k} \| \phi(s) - \phi(x) \| \Big).$$
Here $S$ is the synthetic set and $D_k$ is client $k$'s private data. The term $g_\theta(\cdot)$ is the model update produced by local training from current weights $\theta$, and $\phi$ is a frozen public feature extractor. The margin $\tau$ sets how close a synthetic example may get to a real one, and $\lambda$ weights the penalty.
This is a non-cryptographic, key-free defense, in the spirit of the decoys in DCS². DCS² synthesizes decoy samples that mislead inversion of an ordinary gradient. Here the decoy is itself the transmitted payload.
Relation to Prior Work
Prior work motivating the gap is the synthetic-proxy line itself. FedSynth transmits update-matching synthetic data, and FedDM transmits distribution-matched synthetic data. Both report communication gains and neither measures reconstruction. The attack literature supplies the reason for concern. Concealing Sensitive Samples against Gradient Leakage in Federated Learning argues that inversion succeeds when per-sample gradient contributions are weakly entangled, which is precisely the small-client regime. Gradient Inversion with Generative Image Prior shows that image priors restore reconstructions from larger batches. We reuse such priors as the strongest attacker against compressed-gradient baselines.
Prior work constraining the design comes from evaluations of compression as a defense. Evaluating Gradient Inversion Attacks and Defenses in Federated Learning finds that magnitude pruning alone needs more than 99% sparsity to stop inversion, at about a 10% accuracy cost. Mixed Quantization Enabled Federated Learning to Tackle Gradient Inversion Attacks keeps quantization scales private on the client and drops DLG reconstruction quality from 55.3 dB to 8.1 dB PSNR on MNIST. DLG (Deep Leakage from Gradients) is a standard optimization-based inversion attack. Together these mean the comparison class is not "uncompressed gradients" but compressed gradients that may already be fairly hard to invert. A synthetic proxy must beat those.
The communication side supplies the strongest bandwidth competitor. Practical Hybrid Gradient Compression for Federated Learning Systems combines sparsification, quantization and predictive coding. It reports 15,286× uplink compression with 90.26% CIFAR-10 accuracy using ResNet-18, against 90.52% for uncompressed FedAvg. Any proxy method sending more bits than that needs a privacy argument to justify itself.
We deliberately do not build on parameter-efficient fine-tuning. PEFTLeak shows that adapter sparsity offers little protection against an active server. The FedRot-LoRA rotation symmetry (FedRot-LoRA) might look like a free obfuscation, but we argue below that it is not.
Research Program
The work has three parts.
The first is a leakage audit of FedSynth and FedDM. We sweep the synthetic budget per class, the local dataset size, and the local training steps. We record three exposure measures. The first is direct exposure: for each private example, its best structural similarity (SSIM) and LPIPS distance to any received synthetic example, needing no attack at all. The second is enhanced exposure: the same measure after the server refines the proxy with a generative-prior inversion of the gradient that the proxy implies. The third is membership inference AUC.
The second part is a matched-bits comparison. Every uplink representation is placed on the same axis of bits per client per round. These are the dense update, Top-$k$, HGC-style hybrid compression, private-scale mixed quantization, and the two synthetic proxies. Each is attacked with the strongest applicable attack: DLG-style cosine matching with a generative prior for gradients, and direct plus enhanced exposure for proxies. This yields an exposure-versus-bits curve per representation at fixed final accuracy.
The third part is the repulsion-constrained synthesis above. We tune $\tau$ on a held-out client population. We study whether the margin can be set from public data alone, so that clients need no per-deployment tuning. We also measure the added client compute, since synthesis is already expensive.
The core prediction is falsifiable. Unconstrained gradient-matching proxies should show direct exposure rising sharply as local data shrinks toward the synthetic budget. Compressed gradients at equal bits should not.
Baselines
The naive reference is plain FedAvg with dense uplink updates attacked by DLG with a generative prior. It anchors both the accuracy ceiling and the worst-case gradient exposure. The relevant published undefended figure is 55.3 dB DLG PSNR on MNIST from the mixed-quantization paper, and 59.20 dB undefended PSNR on MNIST reported by DCS².
The strongest communication competitor is HGC-style hybrid compression, reimplemented with its intrinsic, cross-client and temporal stages. It is strongest because it reaches far smaller payloads than any synthetic proxy at near-uncompressed accuracy (the 90.26% CIFAR-10 figure above). It is run at the same number of rounds, with the same client sampling and data partitions as the proxies. Its compression ratio is chosen so that its bits per round match each proxy budget.
The strongest privacy competitor at low cost is private-scale mixed quantization. Its reported 8.1 dB DLG PSNR on MNIST at 4× compression is the number to beat on exposure.
The proxy baselines are FedSynth and FedDM run as published. Lacuna has no reported accuracy number for FedSynth on FEMNIST, MNIST or Reddit, only qualitative superiority over random masking. FedDM is reported only as about 80% per-round communication reduction, so we reproduce both.
A DP baseline adds calibrated Gaussian noise to the synthesis gradients, giving a formally private proxy. It is tuned to match the accuracy of our method.
Two ablations isolate the key ingredient. One initializes synthetic examples from noise instead of real samples, with no repulsion. The other uses repulsion in pixel space rather than feature space.
Every method receives the same hyperparameter search budget, measured in full federated runs, and the same client compute cap per round, measured in local forward-backward passes. Our method beats the field if both of the following hold. Its fraction of private examples exposed above a pre-registered SSIM threshold falls below that of mixed quantization at equal bits. Its final accuracy stays within one point of unconstrained FedSynth.
Why not just…?
Why not just initialize the synthetic data from random noise? This is the most serious threat. It is cheap and may already prevent visual near-copies. We test it directly as an ablation. Our expectation is that noise initialization delays rather than prevents convergence toward real examples when local data are scarce, because the gradient-matching optimum does not depend on where optimization starts. We also expect it to leave membership inference unchanged. If noise initialization alone keeps both exposure measures at the level of our method, the repulsion term is unnecessary. The audit, the first part of the program, would then remain the contribution.
Why not just add differential privacy to synthesis? This gives a formal guarantee, which our margin does not. However, the material reports that DP noise sized to stop inversion dropped MNIST accuracy from 96.67% to roughly 78–85% (Mixed Quantization). Synthetic proxies have little capacity to absorb noise. We include DP synthesis at matched accuracy. If it matches our exposure at equal accuracy, we will recommend it instead, since it comes with a proof.
Why not just send compressed gradients and forget proxies? At the reported 15,286× uplink compression of HGC, proxies may simply lose on bits. They may also lose on privacy, because more than 90% pruning still did not stop inversion in Evaluating Gradient Inversion Attacks while extreme compression might. The matched-bits comparison settles this. If compressed gradients dominate on both axes, the honest conclusion is that synthetic-data uplinks should be retired, and that finding is itself worth publishing.
Why not just hide LoRA updates behind secret rotations? Low-rank adapters are invariant to rotations, which seems to offer free obfuscation. But the server knows the factors it broadcast. It can recover a client's rotation by Procrustes alignment against that reference, the same operation FedRot-LoRA uses for aggregation. Moreover, the product update is unchanged by rotation, so attacks such as PEFTLeak that read adapter structure are not blocked in principle. We therefore do not pursue it.
Proposed Experiments
We use MNIST, FEMNIST with its natural writer partition, and CIFAR-10 under Dirichlet-0.1 label skew. These match the settings where FedSynth, mixed quantization and HGC report results. The decisive experiment is the exposure curve over local dataset size at a fixed synthetic budget. The core claim fails if unconstrained FedSynth shows no rise in direct exposure as clients shrink.
The second check is the matched-bits frontier. Our method must lie strictly inside the region bounded by mixed quantization on exposure and by FedSynth on accuracy.
The third check is the noise-initialization ablation, reported with membership inference AUC as well as SSIM. Visual dissimilarity can hide membership leakage, and only both measures together justify the repulsion term.
We will report client compute alongside all results. Proxy synthesis is already the method's weak point, and a defense that doubles it must earn that cost.
References
- FedSynth: Gradient Compression via Synthetic Data in Federated Learning
- FedDM: Iterative Distribution Matching for Communication-Efficient Federated Learning
- Concealing Sensitive Samples against Gradient Leakage in Federated Learning
- Gradient Inversion with Generative Image Prior
- Evaluating Gradient Inversion Attacks and Defenses in Federated Learning
- Mixed Quantization Enabled Federated Learning to Tackle Gradient Inversion Attacks
- Practical Hybrid Gradient Compression for Federated Learning Systems
- Gradient Inversion Attacks on Parameter-Efficient Fine-Tuning
- FedRot-LoRA: Mitigating Rotational Misalignment in Federated LoRA
- MineGrad: Gradient Inversion Attacks on LoRA Fine-Tuning