In personalized federated learning, phones and hospitals jointly train a shared starting model. Each client can then adapt that model to its own data with one or two gradient steps. The standard way to train such a starting model is meta-learning in the style of MAML. Each client takes an inner step on one local batch, the support batch. It then computes the gradient of the loss on a second batch, the query batch, at the adapted point, and sends that gradient to the server. Exact meta-gradients require second derivatives, so deployed systems use first-order shortcuts. One drops the curvature term entirely; another estimates it from finite differences. Gradient inversion attacks reconstruct a client's training images by searching for inputs whose gradient matches the one the client sent. They have been studied almost entirely for plain SGD and its multi-step variants. Nobody has asked which of a meta-learning client's batches the attacker actually recovers.
This note argues that the answer is structurally lopsided and that it can be predicted before running any attack. To first order in the inner step size $\alpha$, the transmitted first-order update is an ordinary query-batch gradient plus a support-batch signal shrunk by $\alpha$ and filtered through query-batch curvature. Two consequences follow. Query images should be about as exposed as under plain FedAvg. Support images should be exposed only to the extent that a signal-to-noise ratio, scaling with $\alpha$ times query curvature, survives the noise, quantization, or rounding already present in the channel. Which approximation the client uses, pure first-order or Hessian-free, should change support leakage only through one extra, measurable term. We propose to turn this into a role-resolved leakage measure and validate it against the strongest available attack. We then test whether it supports a cheaper, role-aware privacy protocol for meta-learning clients.
Core Claim
Let $\theta$ be the broadcast model, $L_S$ and $L_Q$ the client's support and query losses, $H_S, H_Q$ their Hessians at $\theta$, and $W=\theta-\alpha\nabla L_S(\theta)$ the adapted weights. The pure first-order update of Per-FedAvg (FO) and the Hessian-free update of Per-FedAvg (HF), from Personalized Federated Learning with Theoretical Guarantees: A Model-Agnostic Meta-Learning Approach, expand as follows (a Taylor expansion of $\nabla L_Q$ around $\theta$):
$$g_{FO}=\nabla L_Q(W)\approx \nabla L_Q(\theta)-\alpha\,H_Q\nabla L_S(\theta),\qquad g_{HF}=(I-\alpha H_S)\nabla L_Q(W)\approx g_{FO}-\alpha\,H_S\nabla L_Q(\theta),$$
with errors of order $\alpha^2$.
The leading term depends only on query data. Support data enters only at order $\alpha$, through two channels. In the first-order update it appears only as the support gradient seen through query curvature. The Hessian-free update adds a second channel, support curvature applied to the query gradient, which encodes support inputs differently. Per-FedAvg actually draws the Hessian estimate from a third batch, which then plays the support-curvature role. We therefore predict three things. Query reconstruction quality under an attacker who knows $\theta$ and $\alpha$ will match plain single-gradient leakage. Support reconstruction will fall off with $\alpha\lVert H_Q\rVert$ once any channel noise is present. The Hessian-free update will leak support data measurably more than the first-order one at equal $\alpha$.
We do not claim that support data is safe in a noiseless, infinite-precision channel. There, a scaled-down signal is still exactly recoverable in principle. The claim is that support leakage is noise-limited with an effective signal-to-noise ratio set by $\alpha$, while query leakage is not.
Position Relative to Prior Work
Two papers supply the methods we reuse. Gradient Inversion Attacks Beyond SGD shows that momentum, Adam, and up to eight local SGD steps can be inverted by matching the optimizer's output update rather than a raw gradient. We adopt its update-matching attack by unrolling the client's inner step. That paper never splits recovery by the role a sample plays inside a bi-level update, which is our contribution. Understanding Deep Gradient Leakage via Inversion Influence Functions gives a closed-form leakage measure, called I²F, built from Jacobian-vector products of a single-level gradient with respect to its inputs. It is more than 500× faster than running an attack on ResNet152. We extend it to the meta-update by computing separate Jacobians with respect to support and query inputs.
Two lines of work motivate the question. SoK: On Gradient Leakage in Federated Learning argues that multiple local steps act as a shield and that vulnerability peaks early in training. Our expansion predicts that the inner look-ahead shields the support batch only. Personalized-FL work treats privacy as a byproduct of keeping data local and never audits what the communicated update reveals. This includes Per-FedAvg and Personalized Federated Learning with Server-Side Information, which sends weight differences $(\theta-\phi_i)$ from a proximal local objective.
Theory on first-order meta-learning constrains the analysis. Convergence of First-Order Algorithms for Meta-Learning with Moreau Envelopes shows that first-order meta-learning minimizes a sum of Moreau envelopes, with stationarity neighborhoods of order $\alpha^4$. This tells us how small $\alpha$ can go before utility suffers, which bounds the privacy that can be bought through $\alpha$. Gradient Inversion with Generative Image Prior uses Reptile-style meta-learning offensively, learning an image prior from observed updates. It assumes the victim runs plain SGD; we study the reverse case, where the victim meta-learns.
Research Program
The first component is analytical. We derive role-resolved influence quantities: the singular-value spectra of $\partial g/\partial x_S$ and $\partial g/\partial x_Q$, where $x_S$ and $x_Q$ are support and query inputs. We compute them for the first-order update, the Hessian-free update, exact MAML, and the multi-step weight difference $(\theta-W)$ that Reptile-style and proximal methods transmit. We then compute role-wise I²F scores by modifying the I²F codebase, replacing its single-level gradient with an unrolled meta-update. The analysis predicts a ratio of support to query influence of order $\alpha\lVert H_Q\rVert$ for the first-order update, with an additive support-curvature term for the Hessian-free one. It also predicts that this asymmetry should erode as the number of inner steps grows, because support data then enters at more than order $\alpha$.
The second component is empirical attack validation. We implement update-matching inversion with a differentiable unrolled inner step. The attacker optimizes dummy support and query batches jointly. Labels are recovered with the zero-sum constraint from the Beyond-SGD paper. We report PSNR, SSIM, and LPIPS separately for the two roles. We inject controlled channel noise covering float16 rounding, low-bit quantization, and small Gaussian perturbations. This tests whether support recovery degrades on the predicted $\alpha$-scaled curve.
The third component is a consequence for protocol design. If the asymmetry holds, a client's per-example influence on the update differs sharply by role. Support examples influence it on the order of $\alpha\beta C$, where $\beta$ bounds query curvature and $C$ is the clipping norm. Query examples influence it on the order of $C$. A differentially private client can then calibrate noise to the query batch while giving support examples a tighter bound. To make the curvature bound enforceable rather than assumed, the client clips the support gradient and also clips the realized Hessian-vector product $H_Q\nabla L_S$. The latter is computable at the cost of one extra backward pass. We will report whether this role-aware accounting yields better personalized accuracy than uniform DP at a matched privacy budget. If the curvature clipping costs too much utility, we will say so.
Baselines
The naive reference point is plain FedAvg with one local SGD step on the union of support and query data, attacked by the same update-matching method. It tells us whether meta-learning changes exposure at all. Per-FedAvg reports 58.59% heterogeneous CIFAR-10 accuracy for FedAvg against 71.25% for Per-FedAvg (HF), so we also report the utility cost of falling back.
The strongest existing attack is update matching from Gradient Inversion Attacks Beyond SGD. That paper reports a 3.0 dB PSNR gain over prior SGD attacks, up to 6.5 dB on Adam updates, and partial success up to eight local steps. The file gives no absolute PSNR for it on CIFAR-10. For an absolute anchor at higher resolution, Step-Wise GradInv in High-Fidelity Gradient Inversion in Distributed Learning reports PSNR 15.12 and LPIPS 0.33 on ImageNet with ResNet-18/34 at batch size 8. We will run both attacks with their published regularizers. The attacker gets an identical per-image iteration budget across all update rules and full knowledge of $\theta$, $\alpha$, and batch sizes. This is the honest-but-curious server of the threat models in the file.
The key ablation compares two versions of the attacker. One unrolls the inner step; the other matches the update as if it were a plain gradient at $\theta$. Our expansion predicts that the second recovers query images almost as well, and support images not at all. A second ablation swaps the first-order update for the Hessian-free one at fixed $\alpha$, which isolates the support-curvature channel.
On the defense side, the competitor is uniform DP noise on the whole update at the same privacy budget. We also report the decoy defense of Concealing Sensitive Samples against Gradient Leakage in Federated Learning as an empirical reference. That defense reduces MNIST PSNR from 59.20 to about 7.84 and reaches 65.68% TinyImageNet accuracy. Beating these means one of two things at a matched budget: higher personalized accuracy at equal support- and query-role reconstruction PSNR, or lower PSNR at equal accuracy.
Why not just…?
Why not just run the Beyond-SGD attack on the meta-update and report a single number? That attack is our measuring instrument, but a single pooled PSNR hides the result we care about. If support and query leak very differently, averaging them yields a middling number that neither reassures nor alarms anyone. The unrolled-versus-plain ablation and the role-split metrics are what make the measurement informative. Running the attack alone also gives no way to predict leakage for a new $\alpha$ or approximation without re-attacking. The role-resolved influence scores are meant to fill that gap, and we will test that claim by rank correlation against attack PSNR.
Why not just treat the meta-update as two local SGD steps and rely on the SoK multi-step shield? The expansion says the shield is one-sided. The query term is a direct gradient that no inner step obscures, so the SoK intuition would overstate protection for half the client's data. If our query-role PSNR comes out well below the FedAvg baseline, the multi-step framing was adequate and our central claim is wrong.
Why not just apply uniform DP-SGD to the whole update? This is a real threat. If the curvature clipping needed to bound support influence is as costly as extra noise, role-aware accounting buys nothing, and we will report that outcome. The decisive comparison is the accuracy-at-matched-budget experiment below. The Hessian-free variant also needs its own accounting for the curvature batch, which may erase the gain.
Proposed Experiments and Decisive Checks
The main testbed is heterogeneous CIFAR-10 with the Per-FedAvg partition and a small CNN, which allows direct comparison with Per-FedAvg's accuracy numbers. For scale, we add FEMNIST, partitioned by writer, and 64×64 ImageNet with ResNet-18 to stress resolution, following the GIAS and Step-Wise GradInv settings. Within each, we sweep $\alpha$ over a range spanning the values where FO-MAML still trains well, compare first-order, Hessian-free, and exact meta-updates, and vary the number of inner steps from one to eight. We also test early, middle, and late training rounds, since the SoK reports early-training peaks in vulnerability.
The first decisive check is the slope of support-role PSNR against $\log\alpha$ under fixed channel noise. A clear dependence, alongside a flat query-role curve, confirms the asymmetry. Flat support curves falsify it.
The second is whether the Hessian-free update leaks support images measurably more than the first-order update at equal $\alpha$, as its extra curvature term predicts.
The third is the rank correlation between role-wise influence scores and attack PSNR across all configurations. High correlation would let practitioners audit meta-learning clients at a fraction of attack cost.
The last is the accuracy-at-matched-budget comparison between role-aware and uniform DP. It determines whether the asymmetry is only a diagnostic finding or also yields a usable protocol.
References
- Personalized Federated Learning with Theoretical Guarantees: A Model-Agnostic Meta-Learning Approach
- Gradient Inversion Attacks Beyond SGD
- Understanding Deep Gradient Leakage via Inversion Influence Functions
- SoK: On Gradient Leakage in Federated Learning
- Personalized Federated Learning With Server-Side Information
- Convergence of First-Order Algorithms for Meta-Learning with Moreau Envelopes
- Gradient Inversion with Generative Image Prior
- High-Fidelity Gradient Inversion in Distributed Learning
- Concealing Sensitive Samples against Gradient Leakage in Federated Learning