Deciding in Harmonic Space: Cheap Robust Aggregation for Federated Brain-Graph Models

Hospitals that train a brain-disease classifier together without sharing scans face a dilemma. Plain averaging of their model updates is cheap, but it breaks when sites differ a lot (different scanners, protocols, patients) or when a participant sends corrupted updates. The defenses that fix this compare every client's update against every other client's update before combining them. That comparison costs time proportional to the number of clients squared times the number of model parameters. We propose to make those comparisons in a small coordinate system that every brain model already shares: the vibration modes, or "harmonics," of a standard brain atlas's connectivity graph. The server would decide which clients to trust or group from these short summaries, and would then combine the full updates only once.

The opportunity comes from a structural fact. Brain-network models across sites use the same atlas regions as inputs. Their largest weight matrices are therefore indexed by the same regions, and a fixed basis on those regions is meaningful for every client at once. If the right low-dimensional harmonic summary preserves the distances that separate honest site variation from attacks, robust aggregation could get cheaper and possibly more reliable on heterogeneous neuroimaging data. If it does not, the experiments below will show that cleanly.

Core Claim

Relational aggregation means any server rule that looks at relationships among client updates, such as Krum, geometric median, or nearest-neighbor mixing. For federated brain-graph classifiers, the claim is that the server can split such a rule into two steps. In the first step it makes all pairwise decisions (neighbor sets, selection, robust weights) on harmonic sketches of the updates. In the second step it applies those decisions to the full updates as a single weighted combination.

Let $U_k$ be the first $k$ eigenvectors of the Laplacian of a fixed reference connectome on the shared atlas. The Laplacian is the matrix whose eigenvectors are the graph's smooth-to-rough modes. Let $\Delta W_i$ be client $i$'s update to its region-indexed first layer. The sketch of client $i$ is

$$s_i = \mathrm{vec}\left(U_k^\top \Delta W_i\right),$$

with the edge-indexed version $U_k^\top \Delta M_i U_k$ for models whose inputs are connectivity matrices. Server cost then drops from $O(n^2 d)$ to roughly

$$O\left(n\, d + n^2 k h\right),$$

where $n$ is the number of clients, $d$ the parameter count, and $h$ the layer width. The linear term covers computing the sketches and forming one final weighted sum. The quadratic term covers the pairwise comparisons, which now happen in $k h \ll d$ dimensions.

The second, riskier half of the claim concerns which distances the sketch keeps. Much honest between-site variation, such as global shifts in mean connectivity, may sit in a few specific harmonic bands. If so, comparing clients in a chosen band would reduce the heterogeneity term that limits robust aggregators. It would not merely approximate it.

Relation to Prior Work

The work closest to ours makes the overhead visible but does not reduce it. FedGST and FS²G split a spatio-temporal model between sites and a server. On the server they build a dense subject-level population graph, which is quadratic in subjects, and run an ensemble of graph convolutions over it. Neither uses robust aggregation, and neither uses brain-graph priors on the server. We target the parameter-aggregation step they leave as plain averaging.

Multi-site fMRI Analysis Using Privacy-preserving Federated Learning and Domain Adaptation: ABIDE Results trains an MLP on flattened connectivity across four ABIDE sites with plain averaging and adversarial site alignment. It supplies our naive federated reference point. Personalized Federated Graph-Level Clustering Network computes a client affinity matrix from synthetic graphs that clients produce by expensive min-max training, and it was never tested on connectomes. We avoid both the extra client-side work and the dense full-dimension affinity. Reputation-Based Federated Learning Defense to Mitigate Threats in EEG Signal Classification is cheap but relies on history and ignores graph structure. It serves as a competing low-cost defense.

The method comes from two sources. From Fixing by Mixing we take nearest-neighbor mixing. That method replaces each update with the mean of its $n-f$ closest updates, where $f$ is the assumed number of attackers, before running a robust rule. The authors show it makes standard aggregators near-optimal under heterogeneity, at an explicit $O(d n^2)$ server cost. Our change is to keep their decision logic and move the distance computation into sketch space. From graph signal processing, as used by Sihag et al., we take the harmonic basis. They project fMRI signals onto structural-connectome harmonics and find that low and high graph frequencies carry different, cohort-specific information. This is the motivating evidence that heterogeneity is not spread evenly across harmonics.

Two results constrain the design. Ge et al. show that a globally unified graph topology stabilizes training across subjects, which supports one fixed reference basis over per-site bases. pFedClub shows the cost of relying on a server-side surrogate dataset. Our basis also depends on public data: a reference connectome, for example an HCP group connectome on the same atlas. We will test how sensitive results are to the choice of reference.

Research Program

We will build three pieces. The first is a sketching layer that works on any model whose dominant parameters are indexed by atlas regions or region pairs. This covers a GCN whose first-layer rows are regions, an edge-input MLP in the style of BrainNetMLP, and the region-indexed projections of BioBGT. In these models the first layer holds most of $d$.

The server computes sketches itself from the received updates. Clients do not submit sketches, because an attacker could otherwise send a harmless-looking sketch alongside a harmful update.

The second piece is "decide in sketch, combine in full" variants of three aggregators. For Krum, the server selects one index in sketch space and materializes that one mixed vector. For the geometric median, it runs the iterative reweighting procedure on sketches and applies the final weights to the full vectors. For nearest-neighbor mixing, it chooses neighbor sets in sketch space. Coordinate-wise rules such as trimmed mean cannot be decided in sketch space, so we exclude them and say so.

The third piece is a band study. It asks whether the most useful sketch is the smooth band, a middle band that excludes the constant mode, or the rough band. We will relate the answer to where measured site-to-site variance falls across harmonics.

Baselines

The naive reference is plain federated averaging with no defense, run with the Li et al. site-as-client protocol. That paper reports 67.6% accuracy at the NYU site with Fed-Align, against 60.1% for single-site training. We include it to show how much attacks hurt and how much honest heterogeneity costs.

The strongest existing method is full-dimension nearest-neighbor mixing plus Krum or geometric median from Fixing by Mixing. On extremely heterogeneous MNIST under the "A Little Is Enough" (ALIE) attack, that paper reports about 70% accuracy, against about 10% for vanilla Krum. For this method on ABIDE or ADHD-200 we found no reported result on Lacuna, so we will run it ourselves.

For comparison with the published non-robust federated brain models, we will run the FS²G code. FS²G reports 72.91% accuracy on ABIDE and 72.50% on ADHD-200. FedGST reports 69.53% and 71.27%.

Fairness rules apply across all methods. Every method uses the same backbone, the same graph construction (Spearman correlation via the DataCentricBrainGraphs tools), the same site-wise folds, the same number of rounds, and the same hyperparameter search budget. Any search over $f$ or $k$ counts toward that budget.

The key ablations hold $k$ fixed and replace the harmonic basis with four alternatives. These are a random Gaussian projection, a random orthonormal basis on regions, principal components of the current round's updates, and a mismatched reference connectome. These ablations isolate whether the brain-graph structure matters or only the reduced dimension does.

The thresholds for success are as follows. Beating the strongest baseline means matching full-dimension mixing within one accuracy point under attack while cutting server time by at least an order of magnitude in the many-client regime. Beating it outright on robustness requires a band-selected sketch that exceeds full-dimension mixing under ALIE on held-out sites.

Why not just…?

Why not just use random projections, since they already preserve pairwise distances? This is the main threat. Random projections preserve all pairwise distances up to a small distortion with $k$ growing only logarithmically in $n$, and they need no domain knowledge. If a random projection at the same $k$ matches the harmonic sketch on accuracy and attack robustness, the speed half of our contribution reduces to "random-projection nearest-neighbor mixing." We would report it that way and drop the brain-specific framing. The harmonic basis can win only if it does more than preserve distances, namely if it down-weights the directions where honest sites differ. The band ablation tests exactly that.

Why not just run full nearest-neighbor mixing, since ABIDE has only about twenty sites? At that scale the quadratic cost is genuinely small. We concede this, and we restrict the efficiency claim to regimes where it matters. These include finer atlases, where $d$ grows with the square of the number of regions; many-client settings, such as small groups or individual subjects as clients, or wearable-EEG federations; and the hierarchical edge servers described in the file as compute-limited. At twenty sites the only reason to prefer sketches is the possible robustness gain.

Why not just use reputation scores or principal components of the updates? Reputation scoring as in Zhang et al. costs little per round but cannot stop a client that attacks only once or intermittently. Our single-round ALIE and Mimic attacks probe this weakness. Principal components adapt to the data, but they cost $O(ndk)$ per round, and an attacker can steer them because they are computed from the very updates under attack. We include principal components as an ablation and test this steering directly.

Proposed Experiments

We will run on ABIDE and ADHD-200 with sites as clients. Folds will be leave-sites-out so that robustness is measured on scanners never seen during training. Attacks follow Fixing by Mixing: ALIE, Fall of Empires, sign flipping, label flipping, and Mimic, with attacker fractions of 10–30%. We report balanced accuracy, AUC, and F1, along with server floating-point operations, wall-clock time, and peak memory.

The scaling study splits the data into progressively smaller clients, from 20 up to several hundred. It also varies atlas size from 100 to 400 regions, so we can trace measured server cost against the predicted $O(nd + n^2kh)$ curve.

The band study does two things. It measures the share of between-site update variance in each harmonic band on clean training. It then checks whether the band with the lowest honest variance gives the best robustness. That would be the mechanistic evidence for the second claim.

The decisive check is simple. If harmonic sketches beat random sketches at equal $k$ on held-out-site accuracy under attack, the brain-graph structure is doing real work. If they tie, we have a well-measured efficiency result and a negative result about domain priors.