Fraud detectors built on graph neural networks decide how much to trust each neighbor by comparing two nodes at a time: is this neighbor's feature vector, or its predicted label, close to mine? Camouflaged fraudsters defeat this by design. They choose their own features to look like the benign users they link to. A two-node test only examines quantities the attacker controls. We propose a test the attacker does not control. Learn a small rotation on every edge that says how to translate one node's representation into its neighbor's coordinates. Then multiply these rotations around short closed loops in the graph. On honest parts of the graph the loops should close, so the product returns to the identity. A loop that runs fraudster → benign user A → benign user B → fraudster passes through the A–B edge, which the attacker cannot touch. Making every such loop close requires one feature vector to satisfy a separate constraint for every pair of its benign neighbors. That is impossible once the neighbors are diverse.
This idea is borrowed from how physicists find topological phases of matter. There, a "Wilson loop" composes the step-by-step transport of electronic wavefunctions around a closed path in momentum space. The result does not depend on arbitrary phase conventions, and a non-trivial twist reveals global structure that no single local measurement shows (Topological insulator to Dirac semimetal transition driven by sign change of spin-orbit coupling in thallium nitride). We carry the same quantity onto fraud graphs. We use the per-edge loop twist to reweight neighbors dynamically, and we test whether it resists feature camouflage better than pairwise attention. The test includes an attacker who knows the defense.
Core Claim
The claim is that closed-loop transport inconsistency is a camouflage signal that pairwise neighbor weighting cannot recover. Pairwise filters can be fooled by the attacker's feature choices; loop consistency is far harder to fool because it is overconstrained.
Concretely, attach an orthogonal map $R_{ij} \in O(d)$ to each edge $(i,j)$. Here $d$ is a small dimension (2–4) and $O(d)$ is the set of $d \times d$ rotations and reflections. For a cycle $c = (v_1, \dots, v_k, v_1)$, the holonomy is the ordered product of the maps around the cycle, and the edge inconsistency score averages its deviation from the identity:
$$H_c = R_{v_k v_1} \cdots R_{v_2 v_3} R_{v_1 v_2}, \qquad s_{ij} = \frac{1}{|\mathcal{C}_{ij}|} \sum_{c \in \mathcal{C}_{ij}} \lVert H_c - I \rVert_F .$$
In this definition, $\mathcal{C}_{ij}$ is a sampled set of short cycles (length 3–5) through edge $(i,j)$, $I$ is the identity, and $\lVert\cdot\rVert_F$ is the Frobenius norm. The score $s_{ij}$ multiplies the attention weight of the edge by $\exp(-\beta s_{ij})$, where $\beta$ is a learned scale. We also track a discrete variant, the sign of $\det H_c$. It is $-1$ when the loop composes to a reflection, and it plays the role of a $\mathbb{Z}_2$-style twist index that small feature perturbations cannot change continuously.
What is new is not geometric message passing as such. What is new is computing loop invariants and using them as the trust signal for neighbor weighting in an adversarial setting.
The overconstraint argument gives a falsifiable prediction. A fraudster with $k$ diverse benign neighbors that are themselves connected faces on the order of $k^2$ closure constraints. It has only one $d$-dimensional feature vector with which to satisfy them. Detection power should therefore rise with the number of independent cycles through a node. It should vanish for nodes that attach in a tree-like way, with no cycles through them. We will test this dependence directly rather than assume it.
Relation to Prior Work
Two fraud detectors motivate the gap.
- CARE-GNN (Enhancing Graph Neural Network-based Fraud Detectors against Camouflaged Fraudsters) prunes neighbors by the $l_1$ distance between MLP label predictions. It sets per-relation thresholds with a bandit learner.
- SCSE-GFD (Spectrum-Constrained and Skip-Enhanced Graph Fraud Detection) reweights edges with a supervised homophilic-versus-heterophilic edge classifier and spectral polynomial filters.
Both score edges from the two endpoints alone. The second also depends on edge-type supervision, a weakness its own authors acknowledge. Our score uses no edge labels and looks beyond the two endpoints.
Three lines of work supply the transport machinery, but none uses loops as evidence.
- Neural Sheaf Diffusion (Neural Sheaf Diffusion) learns per-edge maps from features. It shows that orthogonal maps with modest dimension separate classes under heterophily.
- Conn-NSD (Sheaf Neural Networks with Connection Laplacians) fixes the maps by aligning local principal-component bases. Its local 1-hop estimate is exactly what injected edges corrupt.
- GESC (Gauge-Equivariant Graph Networks via Self-Interference Cancellation) uses complex phase rotations per edge with attention-style gating.
All three diffuse along edges and never multiply transports around cycles. None has been tested on fraud graphs or against camouflage.
The physics papers supply the core object, a transport composed around a closed loop, which is independent of how coordinates are chosen at each node. In the TlN and TaIrTe4 work (Dual quantum spin Hall insulator by density-tuned correlations in TaIrTe4), Wilson loops are the standard check precisely when simpler symmetry-based tests are unavailable. We borrow that logic: when pairwise similarity has been engineered to be uninformative, closed-loop quantities are what remain.
Gauge-invariant representation holonomy motivates the claim that loop twist carries information that pointwise similarity misses. That paper finds two image models with near-identical alignment scores (CKA 0.987) but very different holonomy. It also finds that holonomy correlates with adversarial robustness (Pearson $r \approx 0.8$), with adversarially trained models showing higher holonomy.
That finding might seem to cut against our use of low holonomy as "honest." It does not, because the two works measure different objects. They measure the twist of a model's representation map over input space. We measure the closure of an edge connection fit to a particular graph. We take from them only the evidence that loop quantities are not redundant with pairwise ones.
Research Program
Fitting the transports. The transports must not simply follow whatever the attacker's features say, so we fit them in two stages.
- Label-free stage. Fit the maps $R_{ij} = \phi(h_i, h_j)$ by minimizing a robust total loop inconsistency over all sampled cycles. We use a trimmed loss that ignores the worst fraction of loops. Fraud is rare, so the fitted connection should be nearly flat on the benign bulk of the graph, and residual frustration should concentrate on inserted edges. This stage uses no labels, which addresses the label scarcity typical of fraud data.
- Joint stage. Train the inconsistency-gated network on node classification. The loop regularizer is kept to stop the maps from absorbing the camouflage.
Cycles on multi-relational graphs. On YelpChi, cycles are allowed to mix relations. CARE-GNN reports that the "same user" relation is about 90% label-consistent while "same product" is about 5%. Mixed loops therefore tie camouflaged edges to reliable ones. This is the mechanism we expect to matter most, and we will test it with an ablation restricted to single-relation cycles.
Cost. Short-cycle sampling is done once per graph, with a per-edge cap on the number of loops. Holonomy computation then costs $O(|\mathcal{C}| k d^3)$ per epoch and is batched with the forward pass.
Analysis questions. We will test three things:
- whether $s_{ij}$ ranks known fraud–benign edges above benign–benign edges without any supervision;
- whether detection gain scales with the cycle count through a node, as the overconstraint argument predicts;
- whether the discrete determinant-sign invariant survives adaptive attacks better than the continuous score.
Baselines
Naive GNNs. The do-nothing reference is uniform or attention-based aggregation: GCN and GAT (Graph Attention Networks). CARE-GNN reports these collapse to roughly 55% AUC on Yelp. We rerun them on our splits so that all comparisons share data and preprocessing.
CARE-GNN. This is the canonical pairwise camouflage defense. It reports about 75% AUC on Yelp and about 89% on Amazon. We run it from the DGFraud toolbox released with that paper.
SCSE-GFD. This is the strongest existing method on the question. Its reported scores are:
| Dataset | AUC | F1-macro |
|---|---|---|
| YelpChi | 90.58% | 76.12% |
| Amazon | 96.21% | 92.28% |
It is the most direct competitor because it already performs dynamic heterophily-aware edge reweighting. Beating it on clean data is not required. Amazon in particular leaves little headroom. The bar is either of two outcomes:
- matching it within noise on clean data while exceeding it under injected camouflage;
- exceeding it when edge-type supervision is removed or corrupted, which is the regime its authors flag as its weakness.
Geometric baselines without loops. O(d)-NSD and GESC are run on the fraud graphs with the same stalk dimension. Neither has been reported on these datasets; no reported result found on Lacuna for NSD or GESC on YelpChi or Amazon.
Fairness and ablations. All methods get the same training/validation/test splits, the same hyperparameter search budget in trials and GPU-hours, and the same class-imbalance handling. Three ablations isolate the key ingredient:
- Pairwise-only consistency: the same learned maps and gate, but the score uses only the two endpoints' alignment and no loop.
- Cycle counts as features: triangle and 4-cycle counts per edge and per node fed to the strongest baseline, without any transport.
- Random-map holonomy: loop products of untrained random rotations.
If the full model does not beat all three, the loop-transport claim fails.
Why not just…?
Why not just use graph attention, which already weights neighbors dynamically? Attention scores are computed from the two endpoints. An attacker who matches benign features receives high weight. CARE-GNN's reported ~55% Yelp AUC for general GNNs shows how badly this holds up under camouflage. GESC's depth stress test shows GAT collapsing when stacked to 12 layers. GAT is in our baselines on every condition, including the adaptive attack.
Why not just count triangles or short cycles around each node? This is the most serious threat. Fraudsters linking to popular, mutually unconnected benign users close few triangles, so cycle counts alone might expose them. We take this seriously in two ways.
- The cycle-count ablation is fed to SCSE-GFD itself. On clean YelpChi and Amazon, if cycle counts match the full model, we will report that the topology rather than the transport carries the signal.
- Our main synthetic attack has the attacker close triangles through accomplice sybil nodes. This keeps cycle counts benign-looking, while the loop transports still have to close through benign–benign edges. If cycle counts also win there, we will drop the transport mechanism and pursue the cheaper method.
Why not just run sheaf diffusion or GESC on fraud data, since they already have geometric transport? Untested on fraud, these could well be strong. Both are in our baselines. Our pairwise-only ablation uses exactly their style of learned per-edge maps. The difference between that ablation and the full model measures what composing around loops adds, separately from what matrix-valued transport adds.
Why not just train a supervised edge classifier as SCSE-GFD does? Edge labels are often unavailable or themselves camouflaged. We run SCSE-GFD with its edge supervision progressively randomized, at 0%, 20% and 50% label noise, to settle whether the label-free loop score degrades more gracefully.
Proposed Experiments
Datasets. We use YelpChi and Amazon, both with their multi-relational structure and native class imbalance. Metrics are AUC and F1-macro, plus recall at a fixed low false-positive rate, which matters operationally.
Clean data. We first reproduce SCSE-GFD and CARE-GNN on our splits and compare against their published figures. We then report every method on clean data.
Injected camouflage. Using benign nodes from each graph, we inject sybil fraudsters in two settings:
- Non-adaptive: features are set to the mean of the benign neighbors, and edges go to high-degree benign nodes.
- Adaptive white-box: the attacker runs gradient descent on its own features and chooses a fixed number of edges to minimize both the classifier's fraud score and its own loop inconsistency.
No reported result exists for these attack settings on Lacuna, so all numbers will be ours.
Decisive checks. Three outcomes would settle the claim:
- Under the adaptive attack, the full model's AUC drop is smaller than SCSE-GFD's and the pairwise-only ablation's, while the accomplice-triangle attack neutralizes the cycle-count ablation.
- The unsupervised score $s_{ij}$ alone, before any classifier training, ranks injected edges above benign–benign edges well above chance.
- Detection gain grows with the number of independent cycles through attacked nodes and is near zero for tree-attached attackers, as the overconstraint argument predicts.
If the first check fails, the method is not a camouflage defense. If the third fails, our explanation of why it works is wrong even if accuracy improves.
References
- Topological insulator to Dirac semimetal transition driven by sign change of spin-orbit coupling in thallium nitride
- Enhancing Graph Neural Network-based Fraud Detectors against Camouflaged Fraudsters
- Spectrum-Constrained and Skip-Enhanced Graph Fraud Detection: Addressing Heterophily in Fraud Detection with Spectral and Spatial Modeling
- Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs
- Sheaf Neural Networks with Connection Laplacians
- Gauge-Equivariant Graph Networks via Self-Interference Cancellation
- Dual quantum spin Hall insulator by density-tuned correlations in TaIrTe4
- Gauge-invariant representation holonomy
- Graph Attention Networks