Federated Mass-Decorrelated Anomaly Detection for Collider Jets

Searches for new particles at colliders often use anomaly detectors. These are models trained only on known Standard Model "background" processes, and they flag events that look unlike anything seen in training. Such detectors have a known failure mode called mass sculpting. The anomaly score ends up correlated with the jet mass, so cutting on the score carves an artificial bump into the mass spectrum, and that bump can look like a discovery. The best available fix, Nuisance Randomized Distillation (NuRD), assumes one site holds every background sample. It trains on several backgrounds at once and uses an adversarial critic (a second network that tries to predict mass from the learned representation) to strip mass information out. We ask what happens when the backgrounds are split across sites that train together without pooling events. This is federated learning: each site trains locally and shares only model updates or summary statistics. In that setting a site may see only one background type, or only a narrow slice of the mass range.

Our claim is that the adversarial critic, the component that prevents sculpting, is the part that breaks under federation. The fix is to make decorrelation a quantity that can be summed across sites. Each site sends per-mass-bin first and second moments of its learned representation. The server aggregates them into global statistics, and every site then penalizes its representation for depending on mass relative to those global statistics. The same aggregated moments also define the Mahalanobis anomaly score, so one exchanged object serves both as the decorrelation target and as the detector. In digital-twin terms, this is a concrete protocol for the missing step from an isolated self-monitoring twin to "connected cognitive" twins that exchange parametric summaries rather than raw telemetry.

Core Claim

Robust Anomaly Detection for Particle Physics Using Multi-Background Representation Learning does two things. It trains a representation on several known backgrounds (QCD jets and W/Z jets), and it decorrelates that representation from jet mass using importance weighting plus an adversarial critic that penalizes mutual information between representation and mass. On the hls4ml LHC jet dataset this gives AUROC 0.914 and a significance improvement of 2.32 with maximum-logit scoring. With Mahalanobis scoring, the mass-sculpting divergence falls to JSD 0.118, against 0.255 for a VAE baseline. The method is entirely centralized.

We expect it to degrade when backgrounds are partitioned, for two structural reasons. First, a client holding only QCD jets cannot learn the multi-background classification task at all. Second, a client whose events cover only part of the mass range has a critic that literally cannot observe mass dependence outside that range. Averaging such critics with FedAvg (federated averaging, which takes the mean of client weights) does not produce a global critic. Averaged adversaries are also known to be unstable.

Our proposal replaces the critic with an additive statistic. Let $z = f_\theta(x)$ be the representation of jet $x$, with $y$ its background class and $b$ its mass bin. Each client computes count-weighted local sums of $z$ and $z z^\top$ for every pair $(y,b)$. The server adds these into global class-and-bin means $\mu_{y,b}$ and covariances $\Sigma_{y,b}$, and into mass-marginal class statistics $\mu_y$ and $\Sigma_y$. Each client then minimizes its local classification loss plus

$$ \lambda \sum_{y,b} w_{y,b}\left(\lVert \hat\mu^{\text{local}}_{y,b}(\theta) - \mu_y\rVert^2 + \lVert \hat\Sigma^{\text{local}}_{y,b}(\theta) - \Sigma_y\rVert_F^2\right), $$

where $\hat\mu^{\text{local}}_{y,b}$ and $\hat\Sigma^{\text{local}}_{y,b}$ are the client's current moments, $w_{y,b}$ is the global fraction of events in that cell, and $\lambda$ sets the decorrelation strength. This penalty is weaker than mutual-information independence because it controls only two moments. But the Mahalanobis anomaly score depends on the representation only through class means and covariances. If those moments are constant across mass bins, the expected background score is constant across mass, which is the property that matters for sculpting. Residual dependence in higher moments is exactly what the JSD metric measures, so the weakness is tested directly rather than assumed away. NuRD's importance weights require the global mass histogram per class, and that histogram is also a sum of local counts.

Research Program

The work has three parts. The first is a failure study. We partition the hls4ml jet dataset across simulated sites and run NuRD with per-client critics under FedAvg. We measure how AUROC, significance improvement and JSD degrade as the partition becomes more extreme. The partitions range from IID, to Dirichlet-skewed class mixtures, to pathological one-background-per-site splits, to mass-disjoint splits where each site holds a different mass window (a stand-in for sites with different trigger or selection thresholds). The mass-disjoint case is the one we expect to be most damaging.

The second part is the moment-sharing method described above. It includes federated importance weighting and a federated Mahalanobis scorer built from the same aggregated $\Sigma_y$. We will also measure communication cost. The per-round payload is the model update plus $O(|Y|\,|B|\,d^2)$ moment entries for a $d$-dimensional representation, so $d$ must stay modest. We will test Top-$k$ compression of the model update, borrowed from Deep Anomaly Detection for Time-series Data in Industrial IoT: A Communication-Efficient On-device Federated Learning Approach. That paper reports 50% communication reduction and up to 300× gradient compression for federated time-series anomaly detection. We reuse its method, not its domain.

The third part is a short, deliberately limited mapping onto the six-level maturity matrix of Towards an Extensible Model-Based Digital Twin Framework for Space Launch Vehicles. That paper reached Level 5, meaning autonomous confidence-based shutdown in a single installation, and left Level 6 inter-twin exchange unrealized. We do not claim to build a Level 6 twin. The contribution is narrower: a specification of what a detector twin would exchange at that level, namely aggregated representation moments and mass histograms. It also records a caution drawn from the building-construction maturity assessments (Constructing a digital twin maturity assessment framework for the building construction phase…), which found that the analytical layer is usually the weakest part of real twins. The maturity framework is therefore the lens for this study, not its mechanism.

Two neighbouring lines of work bound the scope. RESuM builds offline multi-fidelity surrogates for detector shielding design. Intelligent experiments through real-time AI runs graph-network triggers on FPGAs in under 10 μs with fixed weights. Both are centralized, and neither addresses decorrelation. We treat on-trigger federated adaptation as later work, because the latency constraints in the sPHENIX paper make online weight updates a separate problem.

Baselines

The naive reference point is the VAE anomaly detector reported in the NuRD paper, with AUROC 0.881 and JSD 0.255. We will retrain it both federated (FedAvg) and per-site, which shows how much of any gain comes from multi-background learning at all.

The strongest existing method is centralized NuRD, using the paper's reported numbers: AUROC 0.914 and SI 2.32 for nurd-ml, and JSD 0.118 for nurd-md. We will rerun it ourselves on the same 600,000-jet training split, with the same architecture and representation dimension, and with a hyperparameter search budget matched to the federated methods. It serves as the ceiling. Beating it is not the goal. The goal is to recover it: we count success as a federated method that stays within 0.01 AUROC of our centralized rerun and within 0.03 JSD of it under pathological and mass-disjoint partitions.

The strongest federated competitor is NuRD with per-client adversarial critics averaged by FedAvg. It uses the same number of rounds, local steps, and total client compute as our method. To our knowledge there is no reported result on Lacuna for federated NuRD or for any federated mass-decorrelated anomaly detector, so all federated numbers will be new. Beating this baseline means lower JSD at equal or better AUROC, averaged over five seeds, on the non-IID partitions.

Two ablations isolate the key ingredient. The first keeps federated importance weighting (shared mass histograms) but drops the moment penalty. This tests whether reweighting alone suffices. The second keeps the moment penalty but computes the targets from local statistics only. This tests whether the benefit comes from global aggregation specifically.

Why not just…?

Why not just pool the jets centrally? For a single collaboration working on public-style simulation, pooling is simpler and gives the best model. We treat it as the ceiling, not as a competitor. The federated question matters only where data cannot move, for example across experiments with different data policies, or where raw streams are too large to ship. The industrial federated work above documents this bandwidth argument in a different domain. If our moment-sharing method does not come close to the pooled result, the honest conclusion is that federation costs too much here, and we will report that.

Why not just run NuRD locally and average with FedAvg? This is the obvious port of the existing method, and it is the main threat. If client distributions are only mildly skewed, it may work fine. The failure study is designed to settle this. We predict that on mass-disjoint partitions, its JSD rises toward the VAE's 0.255 while ours stays near the centralized value. If it does not rise, the moment method is unnecessary and we will say so.

Why not just share mass histograms and reweight? Importance weighting removes the dependence between mass and class label, but it does not stop the representation itself from encoding mass. The first ablation measures exactly this gap. If reweighting alone matches the full method on JSD, we will adopt it as the simpler recommendation.

Why not just train a separate detector at each site? A single-background site cannot perform multi-background learning at all. The NuRD paper attributes part of its gain over the VAE to having multiple backgrounds, so per-site models should fall toward VAE-level AUROC. The per-site baseline measures how much of the benefit comes from the federation itself.

Decisive Experiments

All experiments use the hls4ml LHC jet dataset, with QCD and W/Z jets as backgrounds and top-quark jets as the held-out anomaly. Metrics are AUROC, significance improvement, and JSD of the background mass distribution before versus after a fixed-efficiency score cut, plus total bytes communicated.

The first decisive check is the mass-disjoint partition. We use four sites, each holding a quarter of the mass range, with both backgrounds present. This is where local critics are blind by construction. If moment-sharing JSD stays within 0.03 of the centralized rerun while FedAvg-NuRD degrades substantially, the core claim holds.

The second check is the pathological one-class-per-site partition. It tests whether shared class statistics let sites jointly learn a multi-background task that none of them could learn alone.

The third check sweeps the number of mass bins and the representation dimension. This traces the trade-off between decorrelation strength, the higher-moment leakage that JSD exposes, and payload size, both with and without Top-$k$ compression of the model update.

A negative result on the first check would kill the idea. A negative result on the third would restrict its use to small representations.