In federated learning, phones and institutions train a shared model by sending parameter updates to a coordination server instead of sharing raw data. This was sold as privacy-preserving, but it is not: an eavesdropping server can run a gradient inversion attack — an optimization that finds dummy inputs whose gradients match the intercepted update, thereby reconstructing the private training images. Meanwhile, the most widely deployed fix for unstable federated training is FedProx, which adds a proximal term — a quadratic penalty that tethers each client's local model to the broadcast global model, preventing client drift on non-IID data. A recent theoretical result proved that this same tether stabilizes formal differential-privacy bounds, earning the proximal parameter a reputation as a "privacy hero."
This proposal asks whether that reputation is wrong. There is an unresolved tension in the literature: the tether bounds update magnitude (good for privacy accounting), but it also confines the local trajectory to a small, well-behaved region around a model the adversary already knows, and it shifts the local Hessian spectrum away from zero — and the analytical leakage theory of Understanding Deep Gradient Leakage via Inversion Influence Functions says inversion is hardest precisely when that spectrum is near-singular. Both things can be true at once: the proximal term can shrink the noise needed for a formal differential-privacy guarantee while making the raw, un-noised update easier to invert. Nobody has derived the inversion theory for proximal-constrained updates, built an attack that knows about the tether, or measured which effect wins. We propose to do all three, and to turn the answer into a practical rule for jointly setting the tether strength and the noise level.
Core Claim
The claim we set out to test: proximal regularization, deployed for optimization stability, systematically increases the reconstructability of client training data from transmitted updates, and this hazard must be priced into any private federated system that uses it. The mechanism is concrete. A FedProx client locally minimizes
$$h_i(w; w^t) \;=\; F_i(w) + \frac{\mu}{2}\|w - w^t\|^2,$$
where $F_i$ is the client's empirical loss, $w^t$ is the global model the server broadcast at round $t$, and $\mu \ge 0$ is the tether stiffness. At a local optimum $w^\*$ the stationarity condition is $\nabla F_i(w^\*) + \mu(w^\* - w^t) = 0$, and the Jacobian of this map with respect to $w$ becomes
$$\nabla^2 F_i(w^\*) + \mu I,$$
i.e., every eigenvalue of the loss Hessian is shifted up by exactly $\mu$. Two consequences follow. First, the input-to-gradient mapping loses the near-singular directions that make inversion ill-posed — the same directions that Understanding Deep Gradient Leakage via Inversion Influence Functions shows are responsible for the protective effect of Kaiming and Xavier initialization. Second, the tether force $\mu(w - w^t)$ is deterministic and public: the adversary knows $\mu$ and issued $w^t$, so unlike noise it adds zero adversarial uncertainty while straightening the trajectory — and straighter trajectories are exactly what fast linear-surrogate attacks assume.
This claim stands in direct tension with Convergent Differential Privacy Analysis for General Federated Learning, which proves that the proximal term gives a stable lower bound on privacy under $f$-differential privacy and that constant noise suffices across unlimited rounds (their inversion-attack success rates stabilize around 70% under constant noise rather than climbing to 100%). We do not dispute that result; we sharpen it. Their analysis treats inversion as a post-hoc validation check with an attacker who ignores the proximal structure. Our point is that the required noise level $\sigma$ is set by whichever of two opposing effects binds: $\mu$ reduces the sensitivity the noise must cover, but if it simultaneously amplifies pre-noise leakage, the $\sigma$ that actually neutralizes a prox-aware adversary may be larger than the formal accounting suggests. Resolving this determines whether $\mu$ is free privacy insurance or a hidden liability.
What Exists and What Does Not
Four lines of work approach this intersection and stop short. Convergent Differential Privacy Analysis for General Federated Learning proves the convergent-privacy benefit of the tether but never models an adversary that exploits knowledge of $\mu$. Consideration of FedProx in Privacy Protection builds DP-Prox and PDP-Prox, federated systems combining differential privacy with the proximal penalty (gaining up to nearly 5% accuracy on MNIST and Fashion-MNIST under high-privacy constraints), but evaluates privacy purely through $(\epsilon,\delta)$ accounting with no reconstruction attack ever run — so the empirical vulnerability of proximal updates is unmeasured. On the attack side, Surrogate Model Extension (SME): A Fast and Accurate Weight Update Attack on Federated Learning inverts multi-step local updates by optimizing along a linear interpolation between the initial and final local models, and AGIC: Approximate Gradient Inversion Attack on Federated Learning approximates multi-epoch updates as one giant batch — but both explicitly assume unconstrained SGD, and both acknowledge that non-linear update paths break their approximations; the proximal restoring force is a non-linearity neither models. Finally, Gradient Inversion Attacks Beyond SGD supplies the methodological template we reuse — it breaks adaptive optimizers by writing the attacker's matching loss directly in terms of the optimizer's update equations, gaining up to 6.5 dB PSNR on Adam updates — but its update equations contain no proximal term. No paper in the file derives inversion theory, or builds an attack, for explicitly proximal-constrained objectives.
Research Program
The work has three parts that feed each other. The first is theory: we extend the Inversion Influence Function ($I^2F$) — the closed-form analytical metric from Understanding Deep Gradient Leakage via Inversion Influence Functions, which predicts reconstruction risk from Jacobian-vector products without running any attack — to the proximal stationarity mapping above. The key derivation is how the lower bound on reconstruction error, $I_{\text{lb}}$, changes as a function of $\mu$ once the Hessian acquires the $\mu I$ shift. Our working hypothesis, from the singularity argument, is that the bound tightens (reconstruction error falls) monotonically in $\mu$; proving or refuting monotonicity is itself a deliverable, because it gives a cheap, attack-free predictor (the original $I^2F$ computes over 500x faster than iterative attacks) that practitioners can run before deploying FedProx.
The second part is the attack, which we call ProxSME. Following the update-equation-matching recipe of Gradient Inversion Attacks Beyond SGD, the adversary's matching loss simulates local training including the known penalty $\frac{\mu}{2}\|w - w^t\|^2$ while optimizing dummy data — the tether contributes nothing to the adversary's uncertainty since $\mu$ and $w^t$ are public. A cheaper variant adapts SME's surrogate interpolation: instead of a straight line between the initial and final local states, we search along a proximal-corrected path that accounts for the elastic restoring force bending the trajectory. ProxSME yields a falsifiable prediction even without modification: because the tether straightens trajectories, unmodified SME should perform better on FedProx updates than on FedAvg updates, with reconstruction quality rising in $\mu$.
The third part is defense co-design. If the hazard is confirmed, the fix is not to abandon the tether — On the Convergence of Federated Optimization in Heterogeneous Networks shows it buys a 22% average absolute accuracy improvement over FedAvg when 90% of devices are stragglers, and Convergence and Accuracy Trade-Offs in Federated Learning and Meta-Learning shows it permits higher local learning rates without divergence. Instead we characterize the joint $(\mu, \sigma)$ trade-off surface: $\mu$'s sensitivity-bounding lets $\sigma$ shrink per Convergent Differential Privacy Analysis for General Federated Learning, while its leakage amplification (our Part 1) pushes $\sigma$ back up, and clipping bias — which Fundamental Limitations of Favorable Privacy-Utility Guarantees for DP-SGD shows fundamentally distorts the descent direction when the clipping threshold is small — compounds with the over-constraining bias of large $\mu$ already noted by the FedProx authors. We further exploit Understanding Deep Gradient Leakage via Inversion Influence Functions's finding that isotropic Gaussian noise is wasteful and protection comes from noise aligned with the Jacobian's small singular values: inside the proximal trust region the local curvature is explicitly shaped by $\mu$, so the eigenspectrum that directional noise needs is easier to estimate exactly where the tether operates. The deliverable is a rule for choosing $(\mu, \sigma)$ — or, if the frontier is flat, the honest finding that drift should be handled some other way in privacy-sensitive deployments.
Proposed Experiments
We train FedAvg and FedProx under matched conditions on CIFAR-10 and CIFAR-100 (natural images, the standard audit bed of Evaluating Gradient Inversion Attacks and Defenses in Federated Learning and AGIC), FEMNIST (naturally non-IID, partitioned by writer, used by both the FedProx and SME authors), and LFW faces (the high-stakes facial-recognition setting where Gradient-Leakage Resilient Federated Learning showed near-perfect reconstruction of users under server-side defenses), with a Dirichlet non-IID split and a TinyImageNet/ViT-Small configuration for scale as in Convergent Differential Privacy Analysis for General Federated Learning, and the 90%-straggler regime from the FedProx paper so privacy is compared at the operating point where FedProx is actually needed. We sweep $\mu$ at fixed local steps and attack every intercepted update with DLG Deep Leakage from Gradients, SME, AGIC, and ProxSME under a fixed compute budget, measuring PSNR, LPIPS, and MSE of reconstructions against ground truth, and test accuracy for utility. One honesty note: the FedProx paper's own FEMNIST accuracy table is not summarized with numbers in the file — no reported result found on Lacuna for FedProx's FEMNIST accuracy — so we will re-establish that baseline ourselves rather than quote it. Privacy accounting for the defense experiments uses the Moments Accountant tracking of Gradient-Leakage Resilient Federated Learning and the $f$-DP machinery of Convergent Differential Privacy Analysis for General Federated Learning. We also validate the extended $I^2F$ predictor against measured attack PSNR, in the spirit of the original paper's validated linear correlation with ROUGE and BLEU reconstruction scores on WikiText-103 (for which no numeric values are reported on Lacuna, so we establish the image-domain correlation ourselves). Finally, we map the $(\mu, \sigma)$ frontier: for a target $\epsilon$, compare the best achievable accuracy when $\mu$ is co-tuned versus when $\mu$ is fixed at zero and only $\sigma$ varies.
Baselines
The naive reference point is unprotected FedAvg: clients run plain local SGD, transmit raw updates, and we attack them with DLG and SME under the fixed budget. This anchors the leakage scale — DLG reaches an MSE of roughly 0.0069 on CIFAR-100 where earlier methods hover around 0.25 Deep Leakage from Gradients, and SME adds +2.5 to +3.2 PSNR over prior attacks on FEMNIST and CIFAR-100 Surrogate Model Extension (SME) — and it isolates the delta attributable to the tether, since the only difference from our FedProx runs is $\mu > 0$. Utility is matched by training both to within one point of test accuracy before comparing leakage.
The strongest attack competitor is SME, because it is the best method for exactly our setting — multi-step local updates — while staying cheap (under 3GB of GPU memory, up to 100x faster than unrolling attacks, per its authors). We run it unchanged on FedProx updates across the architectures it reports (MLP, LeNet, CNN, ResNet8, ViT) with the same iteration count, optimizer, and memory cap as ProxSME, so any gap is attributable to modeling the tether rather than to search budget. ProxSME counts as winning if it exceeds SME's PSNR on FedProx updates by a clear margin at matched compute, and — separately informative — if unmodified SME itself already scores higher on FedProx than on FedAvg updates, confirming the trajectory-straightening prediction. AGIC AGIC: Approximate Gradient Inversion Attack on Federated Learning, which gains up to 50% PSNR over baselines on multi-epoch FedAvg updates, serves as a second attack reference under the same budget.
The strongest defense competitors are, first, the best heuristic defense on record: GradPrune at 90% sparsity combined with Intra-InstaHide ($k=4$) at batch size 32, which on CIFAR-10 degrades reconstruction to an LPIPS of 0.73 (versus 0.45 undefended) while retaining 86.10% accuracy against a 93.37% non-private baseline Evaluating Gradient Inversion Attacks and Defenses in Federated Learning. Second, the formal-DP representative, server-side DP (Fed-SDP): clip the update, add calibrated Gaussian noise, which the audit in Gradient-Leakage Resilient Federated Learning shows collapses to 87.2% accuracy on MNIST while allowing near-perfect LFW facial reconstruction, against their Fed-CDP (decay) at 97.48% (non-private baseline 97.98%). We run both defenses on top of our FedProx pipelines with identical tuning budgets (the same grid over sparsity, batch size, clipping threshold, and $\sigma$), and our co-designed $(\mu,\sigma)$ directional-noise defense counts as winning if it pushes attacker LPIPS to 0.73 or worse while keeping CIFAR-10 accuracy above 86.10%, and on MNIST clearly above Fed-SDP's 87.2%. Fed-CDP's 97.48% is cited as the per-example-DP ceiling; we do not need to beat it, since it pays a per-example computation overhead our setting avoids.
The key ablation runs ProxSME with the attacker's simulation silently set to plain SGD — the attacker intercepts a FedProx update but models it as unconstrained — and computes our extended $I^2F$ without the $\mu I$ shift. If the prox-aware variants win, the gain comes specifically from modeling the tether; if they do not, the hazard story collapses and we report that honestly.
Why not just…?
Why not just do more local steps or use bigger batches to blur the update? This is the cheapest supposed defense and the literature has already killed it. Surrogate Model Extension (SME) reconstructs data from multi-step local training in under 3GB of memory, and AGIC handles multiple epochs and up to 8 local steps with up to 50% PSNR gains — multi-step averaging moves the effective gradient along the trajectory but does not hide it. Large batches fall to generative-prior attacks Gradient Inversion with Generative Image Prior and to multi-round user-level tracking Beyond Inferring Class Representatives: User-Level Privacy Leakage From Federated Learning, and edge hardware often cannot hold them anyway. Increasing local steps also worsens client drift, costing up to 22% absolute accuracy versus FedProx under heavy stragglers On the Convergence of Federated Optimization in Heterogeneous Networks — so it fails on both axes at once.
Why not just add standard DP noise and be done? This is the most serious alternative, and we take it seriously rather than argue around it. Server-side clip-and-noise (Fed-SDP) fails as deployed: it is a post-processing defense that leaves training-time trajectories exposed and drops MNIST accuracy to 87.2% while allowing near-perfect face reconstruction on LFW Gradient-Leakage Resilient Federated Learning. Per-example Fed-CDP is far better (97.48%) but pays per-example clipping and noise overhead on every training instance, which the same paper flags as a limitation on constrained devices. Most importantly for us: DP-Prox and PDP-Prox Consideration of FedProx in Privacy Protection already combine DP noise with the proximal tether while assuming the tether is benign, and no one has checked that assumption against an attack that knows $\mu$. Our experiments settle it directly: if prox-aware attacks force larger $\sigma$ than the formal accounting implies, current DP-Prox-style systems are under-noised, and our $(\mu,\sigma)$ frontier quantifies by how much. If they do not, we will report that the tether is genuinely free insurance — also a decisive answer.
Why not just prune gradients or obfuscate inputs (InstaHide, MixUp)? Because the price and the guarantee are both wrong. Pruning alone needs more than 99% of gradient coordinates zeroed to stop attacks, costing roughly 10% accuracy on CIFAR-10, and even the best combined defense (GradPrune 0.9 + Intra-InstaHide, LPIPS 0.73) costs 7.27 accuracy points (93.37% down to 86.10%) while providing only computational hardness — an estimated 493 GPU years to break on ImageNet — rather than any formal guarantee Evaluating Gradient Inversion Attacks and Defenses in Federated Learning. These defenses also operate on the update or the input, and so are entirely orthogonal to our question of whether the optimization objective itself leaks; they can be stacked on top of whatever we conclude about $\mu$.
Decisive Checks
Three results decide the proposal. First, the $\mu$-sweep: if reconstruction PSNR rises monotonically with $\mu$ at matched accuracy and fixed attack budget — across datasets, architectures, and all four attacks — the hazard hypothesis is confirmed; if PSNR is flat or falls, the hypothesis is refuted and FedProx is a free defense, which we will report with equal prominence and reconcile against the $\mu I$ spectral argument. Second, the ablation: the prox-aware attack and the shifted $I^2F$ must beat their prox-oblivious twins on FedProx updates, establishing that the effect comes from the tether and not from incidental differences in update magnitude. Third, the frontier: at matched $\epsilon$ and matched test accuracy, co-tuned $(\mu, \sigma)$ with directional noise must dominate the $\sigma$-only frontier; if it does not, the practical conclusion is that privacy-sensitive heterogeneous deployments should trade the tether for a different drift remedy, and our measurements will say exactly how much utility that costs.
References
- Understanding Deep Gradient Leakage via Inversion Influence Functions
- Convergent Differential Privacy Analysis for General Federated Learning
- Consideration of FedProx in Privacy Protection
- Surrogate Model Extension (SME): A Fast and Accurate Weight Update Attack on Federated Learning
- AGIC: Approximate Gradient Inversion Attack on Federated Learning
- Gradient Inversion Attacks Beyond SGD
- On the Convergence of Federated Optimization in Heterogeneous Networks.
- Convergence and Accuracy Trade-Offs in Federated Learning and Meta-Learning
- Fundamental Limitations of Favorable Privacy-Utility Guarantees for DP-SGD
- Evaluating Gradient Inversion Attacks and Defenses in Federated Learning
- Gradient-Leakage Resilient Federated Learning
- Deep Leakage from Gradients
- Gradient Inversion with Generative Image Prior
- Beyond Inferring Class Representatives: User-Level Privacy Leakage From Federated Learning