When Physics Makes Federated Learning Leak: Protecting Electrode Recipes in Multi-Lab CO₂ Reduction Surrogates

Electrochemical CO₂ reduction turns CO₂ and electricity into fuels, but water in the electrolyte keeps diverting electrons into hydrogen gas. This side reaction, hydrogen evolution, is what sets the Faradaic efficiency: the fraction of current that ends up in the product you want. The recipes that suppress it, such as metal ratios, binder polymer loadings and pyrolysis conditions, are exactly what competing labs will not share. A natural fix is federated learning, in which labs jointly train one predictive model by exchanging model updates instead of raw data. The best current models for these electrodes are hybrids: a small neural network maps an electrode's composition to a handful of physical parameters such as porosity and catalyst particle radius, and a differentiable physics model then turns those parameters into predicted product selectivity.

This note argues that this physical structure makes federated training less private than the image-classification settings where multi-round leakage has been studied. Each lab's update passes through a physics model with only a few unknowns, so the question "which recipe produced this gradient?" becomes a small, well-posed inverse problem. Science federations also have few members, typically 5 to 20 labs, which shrinks the crowd that secure aggregation relies on to hide any one participant. We propose to measure this effect, build a defense matched to how automated labs actually run, and verify lab updates cheaply. The verification uses the fact that the physics side is small and mostly arithmetic, which avoids the hours-long costs of general zero-knowledge proofs for machine learning.

Core Claim

In cross-silo federated training of hybrid physics–machine-learning surrogates for competitive hydrogen evolution, proprietary electrode compositions can be recovered from the sequence of shared updates. We expect recovery in far fewer rounds, and at larger local batch sizes, than the same attacks need against unconstrained networks. Two things drive this. The physics layer compresses each update onto a low-dimensional manifold. The small number of labs makes the participation pattern easy to invert. Protection is still possible without destroying the active-learning speedup, by combining three pieces: aggregation grouped to match the labs' experimental schedule, updates restricted to the physical-parameter head, and spot-checked proofs that each update came from data satisfying charge conservation.

The novelty is the first claim. No paper in the file measures multi-round leakage for physics-constrained scientific surrogates. If physics turns out not to accelerate leakage, the rest of the program changes, and we say how below.

What Prior Work Establishes and Leaves Open

The application setting and the model come from A physics-based data-driven model for CO$_2$ gas diffusion electrodes to drive automated laboratories. That work couples a 1D continuum electrode model with a small multi-layer perceptron (a feed-forward network) that infers unmeasured porosity and particle radius from composition. A Gaussian process, a probabilistic regressor that returns calibrated uncertainty, corrects the residual errors. On 90 electrodes from the AdaCarbon automated platform, the combination sped up the search for high ethylene selectivity 3.1x over random sampling. It is entirely centralized. We reuse its model as the object being federated and attacked.

The chemistry explains why small errors matter. Continuous coordination modulation with different heteroatoms unveils favorable single-atom Ni sites for near-unity CO selectivity in CO2 electroreduction shows that swapping one coordinating atom flips a nickel site between 99.3% CO selectivity and hydrogen-dominated behavior. Selectivity is therefore a sharp function of subtle recipe choices. That sharpness makes recipes valuable. It also means the model's gradients are strongly informative about them.

Three results on multi-round leakage motivate the threat model. Securing Secure Aggregation: Mitigating Multi-Round Privacy Leakage in Federated Learning shows that when the set of participants changes between rounds, a curious server can solve a linear system and isolate individual updates within $O(N)$ rounds for $N$ clients. The same paper proposes all-or-nothing batch participation as the fix. Lam et al. formalize the same disaggregation. Temporal Gradient Inversion Attacks with Robust Optimization reconstructs training inputs by combining gradients from many rounds, reaching batch size 128 where single-round attacks fail beyond 8, and it survives differential privacy and 40% sparsification. Feng et al. show that per-round noise averages away when a client trains on a fixed dataset, and Zhu shows the same accumulation for membership signals.

All of this work runs on images and text with many clients. For $N=8$ labs, an $O(N)$ collapse means privacy is gone within a few weeks of experiments. And an attacker that knows the physics model searches a space of perhaps four composition variables rather than a 224×224 image. The disaggregation defense also assumes clients can join rounds together. Automated labs cannot always do that: their cadence is set by synthesis and 24-hour stability runs, and by electrodes that reconstruct under bias (Chemical and Morphological Transformations of a Ag-Cu Nanocatalyst During CO2 Reduction Reaction).

For verification, the file offers a clear cost ladder, and it constrains our design. Full zero-knowledge proofs of machine learning compile the computation into arithmetic over a finite field, and they are slow. Trust from afar reports zkLLM at 92x inference cost, about 10 hours for 64 challenges. Confidential Guardian needs about 333 s per point on ResNet-50 but under 1 s per point on tabular models. zkLLM scales by special-casing non-linear operations. Cheaper routes also exist. Secure and Confidential Certificates of Online Fairness amortizes proofs by checking a random subset of committed queries ("cut-and-choose"), gaining over 1000x. Verifiability and Privacy in Federated Learning through Context-Hiding Multi-Key Homomorphic Authenticators proves that an aggregate is the exact sum of client updates: 2.8 s of client work for a 1M-parameter model and about 10 ms to verify. Its authors note that it does not address multi-round inversion.

Our surrogate head is tiny, so it sits at the cheap, tabular end of this ladder. The continuum solver, which contains exponential electrode kinetics, is where proof costs would blow up, so we deliberately avoid proving it.

Research Program

The attack. We first build a physics-aware multi-round attack. The server sees the aggregated update for each round, meaning the gradient of the hybrid model's loss summed over participating labs. The attack has two stages. Stage one uses the participation matrix to disaggregate individual labs' contributions, following So et al. Stage two runs temporal inversion in the spirit of Li et al., but it optimizes over composition and measured selectivities rather than over pixels, and it pushes each candidate through the known physics model.

The mechanism we test is simple. The shared-network update for a lab equals a sum of per-electrode terms. Each term is a Jacobian (how the network's physical-parameter outputs change with its weights) multiplied by the loss gradient with respect to a physical-parameter vector $z$ of dimension $d$, around 3 to 5. Because $d$ is small and the physics model is fixed and public, every round contributes many equations constraining very few recipe unknowns. We therefore predict recovery error drops sharply after roughly $\lceil m/d \rceil$ informative rounds, where $m$ is the number of recipe variables per electrode. This estimate is simply equation counting, valid when successive rounds' Jacobians are not collinear.

We will compare this against the same attack on a black-box multi-layer perceptron and on a Gaussian process with no physics, trained on identical data.

The defense. It has three parts.

  • Schedule-aligned batch partitioning. We adapt Multi-RoundSecAgg so that each group of labs aggregates only at synchronization points set by their experiment calendars. A lab that misses a point contributes a committed "empty" update rather than dropping out. This keeps every observed sum over at least $T$ labs even when a robot is down.
  • Physical-head-only sharing. Only the network head that predicts physical parameters is federated. Each lab's Gaussian-process residual corrector stays local, which removes the most recipe-specific part of the signal.
  • Checks on individual lab updates. Each lab signs its update with the linear homomorphic authenticators above, so the aggregate is verifiable. Each lab also commits to its per-electrode measurements. At random rounds, cut-and-choose in the style of OATH asks it to prove in zero knowledge two things for a sampled electrode:
  • its reported Faradaic efficiencies (for CO, ethylene and hydrogen) sum to at most one and match the measured charge, which is a linear check;
  • its update equals the head gradient on the committed data.

The continuum solver's output enters that proof as a committed input, not as a proven computation. This keeps proof cost at the scale of a small tabular model.

The precision check. Finite-field proofs and authenticators require fixed-point rounding. Chen et al. show that hydrogen-versus-CO₂ selectivity hinges on small energetic differences, so we directly test whether 4-decimal rounding changes predicted selectivity rankings or the active-learning trajectory. This is cheap, and it determines the fixed-point width.

Baselines

The naive reference point is each lab alone. Every lab runs the Grega et al. hybrid model with active learning on its own data only. This sets the utility floor that federation must beat, measured as acceleration factor for ethylene selectivity plus negative log-likelihood and mean absolute error.

The utility ceiling and privacy floor is centralized pooling: all labs' electrodes in one Grega et al. model. The paper reports an acceleration factor of 3.1x over random sampling on AdaCarbon. We rerun it on our federated splits under the same experiment budget, so the ceiling is measured on identical data.

The strongest existing privacy method is plain Multi-RoundSecAgg with uniform or fairness-aware group selection. It is reported to hold privacy levels $T=3,4,6$ indefinitely, with CIFAR-100 accuracy "virtually indistinguishable" from unprotected training. No result was found on Lacuna for it on scientific regression or cross-silo federations of fewer than 20 clients. We run it with identical labs, rounds and group sizes. Beating it means two things. First, equal or lower recipe-recovery error under the physics-aware attack. Second, at least 90% of its acceleration factor when labs have realistic downtime, where we expect its all-or-nothing rule to stall rounds.

The second strongest competitor is differential-privacy federated training using clipped, noised updates and a standard privacy accountant. It is tuned over noise level with the same search budget as our defense. We compare at matched acceleration factor.

The strongest attack baseline is TGIAs-RO applied unmodified, since Li et al. report it recovering inputs at batch size 128 on ImageNet. It lets us separate how much of the leakage comes from knowing the physics.

Two ablations isolate the key ingredients. The first removes the physics layer from the attacker's model. This tests the core claim that physics accelerates leakage. The second removes each defense component in turn: schedule alignment, head-only sharing, and spot-checks.

Why not just…?

Why not just use differential privacy with a proper accountant? This is the most serious alternative. Feng et al. and Li et al. show that naive per-round noise averages away or is bypassed. A correctly accounted mechanism budgets for composition across rounds, so those results do not refute it outright. The problem is scale. With roughly 90 electrodes and a sharp selectivity landscape, the noise needed over hundreds of rounds may erase the 3.1x speedup. We do not assume this. The matched-utility comparison will show it, and if differential privacy achieves comparable recovery error at comparable acceleration factor, we will report that it suffices and drop our defense's complexity.

Why not just share predictions instead of gradients, as in federated Bayesian optimization? Labs could exchange posterior predictions at proposed compositions and never share weights. This is a real threat to the approach, and we include it as a baseline. Our concern is that the sequence of compositions a lab queries and its predicted selectivities are themselves a multi-round trace. FedMIA (FedMIA) shows that collaborative traces reveal membership across rounds. We will run the same physics-aware attack against this interface. If it leaks less at equal utility, that becomes the recommended design.

Why not just prove everything with a zero-knowledge proof? Proving the full continuum solver, with its exponential kinetics, would land in the regime of hours per proof that Trust from afar and Confidential Guardian report for large models. Proving only the small head plus a linear charge-balance check should sit near the sub-second tabular regime. We will measure both on the same hardware.

Why not just use plain Multi-RoundSecAgg? It is designed for exactly the disaggregation attack. It does nothing, however, against inversion of the group aggregate when a group is a sizable fraction of a small federation, and it stalls when one lab is offline. The downtime experiments will show whether this matters in practice.

Proposed Experiments

We build federations of 5, 10 and 20 simulated labs in two ways. The first uses AdaCarbon compositions split by Cu–Ag ratio and ionomer ratio. The second, larger version uses synthetic labs generated from the Grega et al. physics model with distinct latent porosity and radius, which gives the scale needed to measure leakage curves. Lab downtime follows traces with 24-hour stability-test blocks.

The primary privacy metrics are:

  • composition recovery error;
  • recovery error for the latent physical parameters;
  • membership accuracy for tested compositions;
  • the number of rounds until recovery falls below a threshold, plotted against the number of labs and the local batch size.

Utility is measured by acceleration factor, negative log-likelihood and mean absolute error. Cost is measured as authenticator and proof time per round and bandwidth per round.

As a secondary, out-of-regime check, we repeat the leakage-versus-physics comparison on a small federated graph-neural-network surrogate trained on OC20 hydrogen and CO₂-reduction adsorbates (OC20), split by element across labs. Here the model has no explicit physics layer, which gives a natural contrast.

The decisive check is the physics ablation of the attack. If knowing the physics model does not reduce the rounds needed for recovery by a clear margin over the black-box surrogate at matched data, the core claim fails. In that case the proposal reduces to engineering asynchronous secure aggregation for labs, and we would say so.

The second decisive check is the matched-utility comparison against accounted differential privacy and prediction sharing.

The third is the fixed-point rounding test. If 4-decimal rounding changes the ranking of candidate electrodes by selectivity, the authenticator's fixed-point width must be increased and its cost re-measured before any verification claim is made.