Adaptive Weights as a Privacy Side Channel in Federated Meta-Learning

Federated learning trains a shared model across many devices—a hospital network, a fleet of phones—without centralizing the data. Instead, each client sends its parameter update to a server, which averages them. The catch is that these updates are not anonymous dust: an adversarial server can run gradient inversion, treating an intercepted update as an equation to solve for the private training samples that produced it. Two recent trends independently raise the stakes. Adaptive client weighting has the server score each client's update each round and aggregate with non-uniform weights, so better updates count for more. Federated meta-learning has clients send meta-gradients—updates containing curvature (Hessian) information—so the server can learn a shared "launchpad" initialization from which every client can quickly personalize. Both trends are usually justified by accuracy and robustness, and neither has been privacy-audited; no paper in the collected literature examines their combination.

The proposal here is that these two trends interact badly. If the server publishes or can infer the continuous weights it uses, those weights are extra equations: they turn gradient disaggregation—the problem of pulling one client's update out of an encrypted sum—from a hard combinatorial search into an over-determined linear system. In a meta-learning system this compounds with a second, unexamined exposure: the curvature terms in meta-gradients have never been tested against inversion attacks. We will measure both effects, then design a weighting rule that keeps the personalization benefits of adaptive weighting while leaving the server's inference problem underdetermined and each client's inversion risk flat.

Core Claim

The claim has a negative part and a constructive part. Negatively: in federated meta-learning with adaptive weighting, an honest-but-curious server equipped with the published weight sequence can isolate individual clients' updates in far fewer rounds than existing disaggregation attacks require, and the meta-gradients it recovers may leak more private content than ordinary first-order gradients. Constructively: one can choose aggregation weights under explicit privacy constraints—rank constraints on the weight matrix across rounds, and per-client equalization of inversion risk—while retaining most of the utility gain that motivated adaptive weighting in the meta-learning setting in the first place.

Research Program

The backbone system is Per-FedAvg Personalized Federated Learning with Theoretical Guarantees: A Model-Agnostic Meta-Learning Approach, which trains a shared initialization $\theta$ by optimizing each client's loss after one local gradient step, $\theta_i' = \theta - \alpha \nabla f_i(\theta)$, and which we extend with round-varying aggregation weights:

$$\min_{\theta} \sum_{i=1}^{N} w_i^{(t)} \, f_i\big(\theta - \alpha \nabla f_i(\theta)\big), \qquad \sum_i w_i^{(t)} = 1 .$$

The first work item is the attack. Gradient disaggregation Gradient Disaggregation: Breaking Privacy in Federated Learning by Reconstructing the User Participant Matrix shows that Secure Aggregation—the cryptographic trick where the server only sees a sum—can be undone when the server knows who participated when: it reconstructs a binary participation matrix $P$ from check-in metadata via mixed-integer programming, then separates individual updates. That work assumes $P \in \{0,1\}$. Adaptive weighting replaces it with a continuous, known matrix $A_{t,i} = w_i^{(t)}$, so the server's observation across rounds is simply

$$G_{\text{agg}} = A \, G_{\text{ind}},$$

with $A$ fully observed. The hard combinatorial step disappears; the question becomes one of rank and conditioning of $A$ over a sliding window of rounds. We will extend the released disaggregation codebase gdisag/gradient_disaggregation to continuous and quantized weight matrices and characterize rounds-to-isolation as a function of weight entropy and client count. This is the attack that motivates the defense, and its binary-participation version is also our attack-side baseline.

The second work item is a measurement study of meta-gradient leakage, which no paper in the literature addresses. We compare reconstruction quality for three transmitted quantities: the plain first-order gradient, the first-order meta-gradient $\nabla f_i(\theta_i')$, and the Hessian-free meta-gradient $(I - \alpha \nabla^2 f_i(\theta))\nabla f_i(\theta_i')$ from Per-FedAvg. Two prior findings pull in opposite directions, which is why the study is worth running rather than asserting. Understanding Deep Gradient Leakage via Inversion Influence Functions shows that Kaiming and Xavier initializations protect privacy by inducing singularities in the network's Jacobian, and that perturbing parameters can break this masking; the Hessian-free approximation perturbs exactly at $\theta \pm \delta v$, so we hypothesize higher leakage. Conversely, Concealing Sensitive Samples against Gradient Leakage in Federated Learning shows reconstruction depends on the "low entanglement" of individual sample gradients—cleanly separated subspaces—so the mixing of inner- and outer-loop terms might act as natural obfuscation. Either outcome changes how the defense should be built, and we will audit per-client risk cheaply using the Inversion Influence Function machinery of Understanding Deep Gradient Leakage via Inversion Influence Functions, which is reported to be over 500× faster than running iterative attacks.

The third work item is the defense: leakage-aware weighting. We select weights to maximize expected post-adaptation utility—scoring clients by measured post-adaptation improvement, in the spirit of validation-based weighting—subject to two privacy constraints: the weight matrix over any sliding window must stay rank-deficient enough that the server's linear system remains underdetermined even if weights are observed, and the per-client inversion risk (measured via the Inversion Influence Function) must be equalized across clients rather than concentrated. Concretely this means bounding weight entropy, randomizing weights within a utility-preserving band, and applying per-client calibrated perturbations sized to each client's measured risk. Where server-side scoring is unavoidable, we compare against a variant where scores are computed client-side and combined inside Secure Aggregation, following the client-side validation pattern of Rate-My-LoRA: Efficient and Adaptive Federated Model Tuning for Cardiac MRI Segmentation.

Baselines

The do-nothing reference point is Per-FedAvg with uniform $1/N$ weights and no privacy mechanism, which on heterogeneous CIFAR-10 reports 71.25% test accuracy after one adaptation step for the Hessian-free variant, against 58.59% for FedAvg given the same adaptation budget Personalized Federated Learning with Theoretical Guarantees: A Model-Agnostic Meta-Learning Approach. This is the utility ceiling we must approach and the privacy floor we must improve. The strongest utility-side competitor is FedVal FedVal: Different good or different bad in federated learning, a server-side validation-scoring scheme that maintains convergence with 80% malicious clients on Dirichlet-partitioned CIFAR-10 and lifts FEMNIST minority-label accuracy from 32% to 53%; we will port its scoring onto Per-FedAvg meta-gradients and run it with the same data partitions, compute, and tuning budget, and our method beats it only if it matches its post-adaptation utility while keeping reconstruction quality near the aggregated-update floor of 6.1 dB PSNR reported for CIFAR-10 in Gradient Disaggregation—versus 36.5 dB once disaggregation succeeds. The strongest privacy-side competitor is FedDPA Dynamic Personalized Federated Learning with Adaptive Differential Privacy, which keeps Fisher-sensitive parameters local with adaptive clipping and reports 49.75% on CIFAR-10 and 77.27% on FEMNIST at privacy budget $\epsilon=4$; we will run it at matched attack strength, and beating it means higher post-adaptation accuracy at equal or worse reconstruction quality for the attacker. Finally, two ablations isolate our key ingredient: our weighting with the privacy constraints removed (pure utility weighting, testing what the constraints cost), and the hidden-weights variant described above (testing whether the simple fix suffices).

Why not just…?

Why not just add differential privacy to the gradients? Per-example DP is a proven defense—Fed-CDP reaches 97.48% on MNIST while reducing reconstructed faces to noise Gradient-Leakage Resilient Federated Learning—but it protects gradient content, and the side channel we study is the published weights, which sit outside any DP mechanism on gradients. The disaggregation attack in Gradient Disaggregation succeeds precisely because it exploits non-sensitive metadata; DP-noised gradients, once isolated, are still far easier to attack in aggregate-sensitive ways, and DP carries a real utility cost in personalized settings—FedDPA's 49.75% on CIFAR-10 at $\epsilon=4$ sits more than twenty points below Per-FedAvg's undefended 71.25%, though the setups are not identical. Still, DP is a genuine threat to our defense's necessity, so experiment E3 compares both at matched attack strength rather than arguing around it.

Why not just hide the weights inside Secure Aggregation? This is the cheapest real mitigation and we treat it as a first-class baseline, not a footnote. Honestly, if hidden weights match our constrained-weighting defense on the utility-privacy frontier, the simple fix suffices and our contribution reduces to the attack and the measurement study—we would report that. Two things make us think it will not fully suffice: hiding weights only removes the new side channel and returns the attacker to the binary-metadata attack, which is already enough to disaggregate over hundreds of rounds; and server-side scoring methods like FedVal depend on server-held validation data, so moving scoring client-side changes the method and likely its utility. The hidden-weights ablation in E3 settles this directly.

Why not just run everything on standard FedAvg and drop meta-learning? Meta-learning is where the second, unmeasured exposure lives: no paper in the file audits curvature-carrying meta-gradients against inversion, and the decoy-synthesis defense DCS² Concealing Sensitive Samples against Gradient Leakage in Federated Learning is formulated for single-level risk minimization, so it may not transfer to updates evaluated at post-adaptation parameters. But we are honest about the risk: if E2 finds first-order and meta-gradients leak indistinguishably, the meta-learning component becomes a regime choice rather than a mechanism, and the weighting side-channel result would stand on its own in ordinary FL. The proposal is built so that this outcome, if it occurs, still leaves a complete attack-and-defense story.

Proposed Experiments and Decisive Checks

Experiment E1 extends the released disaggregation codebase gdisag/gradient_disaggregation to continuous and quantized weight matrices on CIFAR-10 and CIFAR-100 with synthetic participation up to 1,024 clients, mirroring the scale where the original attack achieves 100% recovery of the participation matrix. Metrics are rounds-to-isolation, conditioning of the weight matrix, post-disaggregation reconstruction PSNR under GIAS Gradient Inversion with Generative Image Prior (which itself reports 15.58 dB against 13.78 dB for pixel-space attacks), and property-inference AUC, where the original attack reports 0.9+ after disaggregation. Experiment E2 runs the meta-gradient vulnerability study on MNIST, CIFAR-10, and LFW faces, comparing plain, first-order, and Hessian-free meta-gradients under both GIAS and the Inversion Influence lower bound, across uniform versus Kaiming initialization to test the singularity-protection hypothesis. Experiment E3 evaluates the defense on Dirichlet-partitioned CIFAR-10 and FEMNIST against all baselines above, reporting post-adaptation accuracy, per-client inversion risk, and reconstruction PSNR under the combined disaggregation-plus-GIAS attack; success means staying within about two points of the best utility weighting while the attack's PSNR stays near the 6.1 dB aggregated floor and per-client risk spread shrinks. Experiment E4 probes the fairness-privacy tension directly: on FEMNIST, we boost minority-label clients as FedVal does (32% to 53%) and measure whether their reconstruction fidelity rises in proportion—a result that would show performance fairness and privacy can pull in opposite directions under adaptive weighting.

The decisive checks are stated in advance. If continuous weights do not reduce rounds-to-isolation relative to binary participation, the core attack claim fails and we will report it. If first-order and Hessian-free meta-gradients produce indistinguishable leakage curves, the curvature story reduces to a negative result, which we will also report. The proposal is built so that each of its three parts—the attack, the measurement, and the defense—can succeed or fail on its own evidence.